AI training data and image licensing

Why raw files and film originals matter to image-model developers

Heeled boots with spotted and striped patterns arranged against bright yellow.
Animal-pattern ankle boots.Photograph by Robert Harper. Robert Harper Photograph Library, Archive.London. From the archive, shown for illustration.

A web JPEG is the last step in a chain of decisions that cannot be undone. Raw files and film sit at the start of that chain, which gives a developer more room to work.

An 8-bit JPEG records 256 levels of brightness in each colour channel. A 14-bit raw file records 16,384. That gap, a factor of 64, is the simplest way to see what separates a camera original from the copy that ends up on a web page, and why the people who build image models care which one they are given.

This article is general information, not legal advice. Checked on 17 November 2026. File quality does not change the legal questions around training on photographs, some of which are decided and some pending; those are followed in the AI copyright cases tracker.

What disappears between the camera and the web?

A photograph published online has usually been through most of the following, in order:

  1. Rendering. The sensor data is converted to a viewable image, with white balance, contrast and colour fixed.
  2. Reduction to 8 bits. Most of the tonal information is discarded.
  3. Retouching and sharpening, tuned for one output size.
  4. Resizing, often to a small fraction of the original dimensions.
  5. Lossy compression, which removes fine detail the eye is less likely to miss.
  6. Re-compression by each platform that handles the file afterwards.
  7. Metadata stripping by some of those platforms, removing caption, creator and technical data.

Every step is a decision made for a human viewer on a screen, and none can be reversed from the result. A scraped collection contains millions of such files, each processed differently by different hands.

What is in a raw file?

A raw file is the camera's record of what the sensor measured, before step one. Nikon's NEF format, like other makers' raw formats, typically stores 12 or 14 bits per photosite, along with the camera's settings as instructions that have not yet been applied.

For a developer that has four consequences.

  • Uniform rendering. A whole collection can be processed through one pipeline, to one colour space and one tone response, so the model sees consistent data.
  • Any size. Files can be rendered at full resolution, or reduced cleanly from the original instead of from an already compressed copy.
  • Headroom. Highlight and shadow detail that an 8-bit rendering clipped is still present.
  • Technical ground truth. The file records camera, lens, focal length, aperture, shutter speed and date as written at the moment of exposure. The metadata viewer shows these fields where a file carries them.

Raw files are also demanding. They are large, each maker's format needs its own decoder, and someone has to make the rendering decisions the camera would otherwise have made. Long-term handling of proprietary raw formats is its own subject, covered in TIFF, DNG and JPEG 2000 compared.

What do film originals add?

A transparency or negative is the physical first generation. Nothing has been rendered, compressed or resized, because nothing digital has happened to it at all. Its digital life begins at the scanner, and the party commissioning the scan chooses the specification.

Medium format matters here because of area. A 6 x 7 frame on 120 film is about 56 by 70 mm, roughly four and a half times the area of a 35mm frame, which is why studios used it for work that had to survive enlargement, as explained in why studios used medium-format transparency film.

The arithmetic of scanning is straightforward:

Scan resolution Pixel dimensions of a 56 x 70 mm frame Megapixels
1,200 ppi about 2,650 x 3,310 about 9
2,400 ppi about 5,290 x 6,610 about 35
4,000 ppi about 8,820 x 11,020 about 97

These are pixel counts, not promises of detail. How much real information a scan holds depends on the film stock, the taking lens, focus, the condition of the original and the optics of the scanner. Past a certain point more pixels record only grain. Choosing that point is part of a scanning specification, discussed in how to scan medium-format transparencies.

Film also carries the colour and tonal character of its emulsion, which differs from a digital sensor's. Whether that is useful depends on the task. It is a property a buyer can assess on samples, not a performance claim.

Why does this matter for training image models?

Resolution is a one-way street. Models have moved towards larger outputs; the SDXL paper of 2023 describes a design aimed at high-resolution synthesis and trained across multiple aspect ratios. A developer can always reduce a large original to the size a training run needs, and can take several different crops from it. Enlarging a small file means generating detail that was never recorded.

Artefacts are learnable. A model learns recurring patterns, including the blocking and ringing left by heavy compression. Clean sources remove one way for damage to be learned as content.

Consistency helps. One rendering pipeline applied to originals gives a dataset without the accidental variety of thousands of different export settings.

Some tasks need the upstream data. Work on denoising, colour reconstruction, high dynamic range and restoration is done on raw or high-bit-depth material, because the JPEG has already discarded what the task is about.

Originals are evidence. A raw file with intact camera data, or a strip of film in a sleeve, is strong support for a claim about who made a photograph and when. That feeds the documentation a buyer needs, described in image provenance for AI. Material made on film is also camera-made by nature, a point that matters more as generated images spread, as the piece on synthetic data versus licensed photographs explains.

Four kinds of source compared

Web JPEG Master TIFF Camera raw Film original
Bit depth 8 8 or 16 12 or 14, typically Set by the scan
Compression Lossy None or lossless None or lossless, typically None until scanned
Processing fixed? Yes Yes No No
Maximum size As published As saved Sensor resolution Set by the scan
Capture metadata Often stripped Usually present Present On sleeves and notes, not in a file
Work needed before use None Little Rendering Scanning, cleaning, colour management

What should a delivery specification say?

Whichever source is used, both sides should agree in writing:

  • the source of each file (raw, scan from original film, scan from duplicate, derivative);
  • pixel dimensions and bit depth;
  • colour space and whether a profile is embedded;
  • file format and any compression;
  • for scans, the scanner type, resolution, and what cleaning or correction was applied;
  • which metadata fields are delivered;
  • a hash for each file, so the delivery can be verified. The image hash and fingerprint tool shows what such a value looks like.

A supplier should state plainly where it cannot yet meet a specification. Unscanned film is the obvious case: the quality is potential until the scan exists.

What the Robert Harper archive holds

The inventory of the Robert Harper Photograph Library records 42,803 digital photographs, held as Nikon NEF raw files, TIFFs and JPEGs, and 52,684 analogue photographs across 134,340 frames, principally 6 x 7 colour transparencies, with black-and-white negatives, 35mm slides, larger-format film and 4,005 Polaroid prints. These counts come from the archive's inventory; they have not been checked frame by frame against the physical material. One folder of 842 transparencies has been scanned. The rest of the analogue material has not, so no claim is made here about scan quality; that will depend on the specification agreed with a scanning partner, and Archive.London is seeking such partners through its digitisation partners page.

The folder-level breakdown is in the public archive inventory and the summary in the dataset card. The archive is owned by 825 Ventures Ltd, the owner and rights holder, and licences are negotiated directly with the single rights holder, subject to rights review. Developers who want to discuss originals, including a first look at film as it is digitised, can start at the AI training data licensing page.

Frequently asked questions

Do AI models need high-resolution training images?

It depends on the model, but the direction has been upward. Early latent diffusion models were trained at around 512 pixels; SDXL, described in a 2023 paper, was designed for high-resolution synthesis across multiple aspect ratios. High-resolution sources can always be reduced to suit a training run. Small files cannot be enlarged without inventing detail.

Why are raw files better than JPEGs for training data?

A raw file holds the sensor's data before white balance, tone curve, sharpening and lossy compression are applied, usually at 12 or 14 bits per channel. A developer can render every file the same way, at any size and in any colour space. A JPEG has those choices fixed and has discarded information to save space.

Are film scans useful for training AI?

They can be. Film originals are first-generation records with no digital processing history, and medium-format frames hold a great deal of detail. Their usefulness depends on how they are scanned: resolution, bit depth, colour management and cleaning all matter. A poor scan of a fine transparency is still a poor file.

How many megapixels is a 6x7 transparency?

Film has no fixed pixel count. A 6 x 7 frame is about 56 by 70 mm, so a scan at 2,400 pixels per inch produces roughly 5,290 by 6,610 pixels, about 35 megapixels, and at 4,000 pixels per inch about 97 megapixels. How much genuine detail those pixels hold depends on the film stock, the lens, focus and the scanner.

Do compression artefacts in training images matter?

They can. A model learns whatever patterns recur in its data, and heavy JPEG compression leaves recurring patterns: blocks, ringing around edges and smeared fine texture. Files that have been saved and re-saved several times carry more of them. Clean sources reduce the chance that a model reproduces compression damage as if it were part of the picture.

Sources

This article is general information, not legal advice. It reflects the position as we understood it on the date shown and may be out of date. Published by Archive.London. Corrections are welcome under our editorial policy.

Licensing the archive for AI and data projects

95,487 photographs from one photographer and one rights holder. Read the dataset card, search the 503-folder inventory, discuss a sample set under NDA, or arrange a first look as scanning proceeds.

See AI and data licensingWrite to the archive