Type the name of a discontinued slide film into an image generator and you get a mood board of other people's scans. The chemistry tells a different story, and one of the favourite effects belongs to the wrong film.
A film look prompt does not summon a film; it summons the average of every scan that somebody once captioned with the film's name.
An image generator asked for "a street scene, shot on Kodachrome" has no experience of Kodachrome. It has been shown digital files that somebody, somewhere, captioned with the word. That is the angle of this piece, and it explains nearly everything the results get wrong: the model is imitating the reputation of a film, assembled from scans of varying honesty, and one of the effects most often requested alongside it, the glowing red halo, belongs to a different kind of film altogether.
We hold a view on this because the archive we look after is mostly film. So what follows is a comparison by reasoning, effect by effect, between what the physical material did and what a prompt can ask for. It is not a test of any particular generator, and it makes no claim about one.
What was Kodachrome, and why can nobody shoot it now?
Kodachrome was introduced in 1935 and retired on 22 June 2009. Minnesota Public Radio's report of the announcement recorded that the film by then accounted for "just a fraction of 1 percent" of Kodak's still-picture film sales, and that a single laboratory, Dwayne's Photo in Parsons, Kansas, was still developing it. Dwayne's stopped accepting film on 30 December 2010 and ran its last roll on 18 January 2011.
The reason one laboratory mattered so much is the reason the film looked the way it did. Almost every other colour film carries its dye-forming chemistry, the colour couplers, inside the emulsion layers. Kodachrome carried none. It was in effect three black-and-white emulsions stacked on one base, each sensitive to a different part of the spectrum, and the dyes were introduced during processing, one colour at a time. The K-14 process that did this ran to 17 steps, beginning with the removal of a black backing layer and ending with drying. Nobody could do it in a bathroom.
That is the first thing a prompt cannot reach. "Kodachrome" is the name of a process as much as a product, and the process is over. A conservator's timeline of the film lists the stocks most people mean when they use the word: Kodachrome II from 1961 to 1974, Kodachrome 25 from 1974 to 2001, Kodachrome 64 from 1974, with the professional version of the 64 surviving until 2009. Those are different films with different speeds, and a prompt flattens them into one adjective.
What does an image model actually learn when it learns a film look?
Large image models are trained on pictures paired with text. LAION-5B, the best documented public example, contains 5.85 billion image and text pairs gathered from the web. Whatever such a model knows about Kodachrome, it knows from files whose caption, alt text or surrounding page used the word.
Consider what stands between a transparency and one of those files. Somebody scanned it, on a machine whose light source and software made their own decisions about colour. Somebody may have corrected it, perhaps to match the slide on a lightbox, perhaps to match a memory, perhaps to match what the internet has agreed Kodachrome looks like. It was converted to 8-bit sRGB, compressed and resized. Then it was captioned, by a person who may or may not have known which film was in the camera. Many pictures tagged with the film's name online are digital photographs wearing a preset.
So the model's "Kodachrome" is a consensus. A consensus has no ground truth, and it feeds itself: every generated picture labelled with the film's name is another vote for the average.
How does real film grain differ from the grain a generator adds?
Grain is not a layer on top of a film image. It is the image. In a black-and-white negative the picture is made of developed silver; in a colour film the silver is removed during processing and what remains is a cloud of dye wherever a silver grain once was. The texture therefore changes with the picture: it differs between thin and dense areas, between one colour layer and the next, and between an exposure that was right and one that was not.
Format changes it too. A 6 x 7 transparency has roughly four and a half times the area of a 35mm frame, so the same emulsion enlarged to the same print size shows far less grain. That is why studio product photography on medium format looks smooth and a 35mm concert frame from the same year looks gritty, a point we take further in why studios used medium-format transparency film.
Applied grain, whether added by a generator or by an editor, is usually uniform noise at one size across the whole frame. It can be pleasant. Our own film grain tool adds exactly that and says so: monochrome grain with an amount and a size, an emulation of nothing in particular. The mistake is to treat the overlay as evidence of a film.
What is halation, and did Kodachrome have it?
Halation is the soft glow around a bright light on film. The film-emulation developer Dehancer gives a clear account of the physics: light passes right through the emulsion, scatters from the back of the film or the inside of the camera, and returns to expose the film from behind. The red-sensitive layer sits deepest, nearest the base, so it takes most of the reflected light. Hence a halo that is red, shading to orange when the reflection is strong enough to reach the green-sensitive layer as well. The same source notes that the halo is larger relative to the frame on small formats and is strongest around tungsten lamps.
Film makers regarded this as a fault and built layers to stop it. Kodachrome's was a black rem-jet backing on the base, and the first step of K-14 existed to soften and wash it off after it had done its job. In other words, the film was engineered to suppress the very glow that film-look prompts like to pile on. Heavy red halos are a signature of motion picture negative under bright practical lights, and they became fashionable in stills through emulation software. Ask for Kodachrome and halation together and you are asking for one film's virtues and another film's faults.
Film original, scan and generated image: what is each one?
| Property | Transparency | Scan of the transparency | Generated film look |
|---|---|---|---|
| What made the colour | Dyes formed in processing, specific to the stock and the process | The scanner's sensor, light source and profile reading those dyes | Statistics of captioned digital images |
| Grain | The image-forming structure, varying with density and format | A sample of that structure at the scanning resolution | Texture synthesised or overlaid, usually uniform |
| Halation | Present or suppressed according to the film's design | Recorded as it occurred | Added as a style cue, often regardless of stock |
| Can it be checked? | Yes: it is the reference | Yes: against the original | Only against other pictures of unknown origin |
| Does it record a scene? | Yes, one exposure at one time | Yes, at one remove | No |
The table is the argument in small. A scan can be wrong, but it can be wrong about something. A generated film look has nothing to be wrong about.
Why do film originals remain the reference for the look?
Because colour is a measurement before it is a taste. Anyone who wants to model what a film did, whether a colourist building an emulation, a researcher evaluating a model, or a restorer deciding how a faded slide once looked, needs the physical object and the conditions under which it is read. The object does not stay still either. Dyes fade, at different rates in the dark and under projection, which we set out in why colour slides fade. A reference therefore needs dated scans, a description of the scanning chain and, ideally, the stock and process recorded per frame.
This is also where a film archive is useful to people working on image models, and where candour is needed about what it can offer. A model that has learned "film" from the open web has learned a caricature. Correcting a caricature requires examples with known answers: this frame is a 6 x 7 transparency, this one a 35mm slide, this one a Polaroid print, this one a raw file from a digital camera. Labels of that kind are a cataloguing task, and they are worth more than any quantity of pictures captioned "vintage".
What does the Robert Harper archive hold that is relevant?
By its inventory count the Robert Harper Photograph Library holds 95,487 unique photographs, of which 52,684 are analogue, and 134,340 analogue frames once variants are included. The principal analogue format is the 6 x 7 colour transparency, alongside black-and-white negatives, contact sheets, 35mm slides and 4,005 Polaroid prints; the digital side consists of Nikon NEF raw files, TIFFs and JPEGs. More than half of the archive is still on film and has yet to be scanned. The full list of 503 folders is published in the archive inventory, and the film holdings are described on the analogue collection page.
Two limits belong here. The inventory records formats, not film stocks, so we cannot tell you today which frames are on which emulsion, and we make no claim that any particular stock is present. That information will come from the film edges and mounts as cataloguing proceeds. And the counts come from the inventory and have not yet been reconciled frame by frame.
What the archive does offer is one photographer's work across several physical processes over about thirty-five years, with the originals in hand. The dataset card states what is known and unknown about it in the form a technical reader expects. If you are working on film emulation, evaluation or restoration and want material whose physical source can be stated, licensing for AI and data use is available by negotiation with the single rights holder, subject to rights review.
Frequently asked questions
Can AI really reproduce the Kodachrome look?
It can produce a convincing pastiche of how Kodachrome scans tend to look online: saturated reds, deep shadows, a warm cast. It cannot reproduce the process, because the model has only seen digital files that somebody labelled Kodachrome. The result is an average of scanners, editing choices and captions, with no way to check it against an original transparency.
What is halation in film photography?
Halation is a glow around bright highlights caused by light passing through the emulsion, reflecting from the back of the film or the camera, and exposing the film a second time. Because the red-sensitive layer sits deepest, the halo is usually red or orange. Anti-halation layers exist to suppress it, so strong halation indicates a film used without one.
Why was Kodachrome different from other slide films?
Most colour films carry dye couplers inside each emulsion layer. Kodachrome did not. Its three layers were essentially black-and-white emulsions, and the cyan, magenta and yellow dyes were introduced one colour at a time during the K-14 process. That made processing complex and centralised, and it ended when Kodak withdrew the chemistry.
Is film grain the same as digital noise?
No. Grain is the visible texture of the image-forming material itself, silver particles in black-and-white film and dye clouds in colour film, and it varies with density, exposure and format. Digital noise is an error added to a signal. A grain overlay applied evenly across a picture is a decoration, which is why it tends to look pasted on.
Why do film originals matter if the look can be generated?
Because a look without a reference drifts. A transparency is a fixed physical record of what one film did with one scene under one light, and it can be re-scanned as often as needed. Anyone building, testing or correcting a film emulation needs originals to measure against, and descriptions of what those originals are.
Sources
- Minnesota Public Radio News: Kodak retiring Kodachrome film (22 June 2009)
- Wikipedia: K-14 process (steps, rem-jet backing, final processing dates)
- Gawain Weaver Art Conservation: Kodachrome timeline (film types, processes and mount designs by date)
- Dehancer: Halation and its simulation (physical cause, layer order, format dependence)
- Schuhmann et al., LAION-5B: an open large-scale dataset for training next generation image-text models (2022)
Published by Archive.London. Corrections are welcome under our editorial policy.

