Generic AI imagery is not a prompting problem. It is a reference problem. When you start at the prompt with nothing behind it, the model supplies the reference, and the model has the same reference as everybody else. The brands that will make distinctive AI work are the ones that organize their own archive into something a person can scroll, cross-reference, and write from. That is an infrastructure project, not a software purchase.
THE STYLIST WHO CAME BACK
From the mid-1980s on, Simon Foxton helped build the visual language of British menswear. He was a fixture at i-D, Nick Knight's go-to menswear stylist for decades, and a regular in The Face and Arena Homme+, with commercial work for Levi's and Stone Island. His 1991 shoot with photographer Jason Evans, made on suburban roads near his home in Ealing, has been acquired by both the Tate and the Victoria and Albert Museum. He found Edward Enninful on the Underground at Hammersmith and put him in front of a camera.
He is retired now. For the past two years he has been making AI images in Midjourney, from a suburban house in West London, for no reason other than enjoyment. Aperture published a profile of the work in May by Phil Bicker.
I am not writing about it because a famous stylist uses AI. Plenty of famous people use AI and most of them have nothing useful to say about it. I am writing about it because his method is backwards from how everyone is being taught. His is the one that works.
THE METHOD
Foxton has kept scrapbooks his entire career. Torn-out magazine pages, collected and combined, going back forty years. In his words, a magpie approach: taking pieces from unrelated sources and putting them together in ways that had not been tried.
That habit did not stop when he moved to AI. It became the process.
He starts by scrolling his own archive. He looks for correlations between images that have no obvious reason to sit together. He finds a story in the overlap. He distils that story into words. Then, and only then, he writes a prompt.
The prompt is the last step and the shortest one.
Compare that to how AI image work usually gets taught inside a brand. Somebody opens the tool. Somebody types a description of the picture they want. Somebody regenerates eleven times because it looks a bit generic. Somebody eventually blames the model.
The model is not the problem. Forty years of collected reference versus a blank prompt box was never a fair fight.
CASTING POLAROIDS, NOT CAMPAIGN IMAGES
Look at what Foxton makes and you notice something else. He picked the least glamorous formats available.
Casting Polaroids. Outtakes. Behind-the-scenes frames. Straight-up portraits, in the tradition of the i-D street shots he was making in the 1980s. He describes wanting the images to read as plausible, and he is fully aware of the irony in using the word about something entirely manufactured.
This is a real craft decision and apparel brands consistently get it backwards. The default instinct is to point AI at the hero campaign image. That is the hardest register to fake and the one audiences examine most closely, and a small error there reads as contempt rather than as a mistake. Documentary registers are more forgiving because they are supposed to be imperfect. Slightly wrong framing is a feature. Available light covers a multitude of sins.
He also tags the work with an AI disclaimer, and much of it is attributed to a fictional photographer called Boss Tweed, a name he used earlier in his career for jobs he preferred not to be credited on. The labelling costs him nothing. Disclosure is only expensive when the work is not good enough to survive it.
YOUR BRAND ALREADY OWNS THE RAW MATERIAL
This is where it stops being about Foxton.
Every apparel brand of any age owns the equivalent of Foxton's scrapbooks. Archive lookbooks. Campaign contact sheets, including everything that did not get selected. Fabric libraries. Trim boards. Seasonal mood decks going back years. Store photography. Fit session images. The founder's original reference folder.
Almost none of it is organized in a way that a person could sit down and scroll through, and almost none of it is tagged in a way that would let anyone find a correlation. It is stored by season and by style number, because it was filed by whoever needed to find one specific thing once.
Season and SKU are storage logic. They are not creative logic. Nobody has ever had an idea by browsing a folder called SS23_FINAL_v4.
| TAGGED THIS WAY | YOU CAN FIND | YOU CANNOT FIND |
|---|---|---|
| Season and style number | One specific product image | Anything you were not already looking for. |
| Mood, era, subculture, palette, material, register | The same product image | Nothing. This is the version that produces ideas. |
WHAT THIS LOOKS LIKE AS A PROJECT
I want to be careful here, because there is a version of this advice that turns into a nine-month digital asset management procurement and produces no creative work at all. That is not the project.
The project is smaller and it is creative-led. Pick a defined slice of your archive. Digitize what is not digital. Tag it along axes a creative person would search on rather than axes an operations person would file on. Then build the scroll-and-correlate step into your concept process as a real meeting with a real output, before anybody opens a generator.
The tagging axes that matter are mood, era, subculture and cultural reference, palette, material and surface, and photographic register. Six is enough. You are building something to browse, not a taxonomy to admire.
FIVE MOVES
- Find out who owns your archive. In most companies the honest answer is nobody, or one person who has been there eleven years and is not in the room when AI gets discussed.
- Digitize one slice, not everything. One decade, one category, or one founder-era folder. Prove the process produces better work before you scope the full project.
- Tag on creative axes. Mood, era, subculture, palette, material, register. Resist the urge to add twenty more fields because it feels more thorough. It is not more thorough. It is unusable.
- Make the correlation step a meeting. Scroll, find two connections nobody would expect, write one sentence of world for each. That sentence is the brief. The prompt comes after.
- Choose the register deliberately. Decide which photographic mode a piece of AI work should sit in before generating. If the honest answer is hero campaign, ask whether AI is the right tool for it at all.
An archive that produces ideas is a creative system, not a storage system, and the difference is entirely in how it is structured and who is allowed to touch it. Flume and Field helps apparel teams turn brand history into a working reference layer and build the concept ritual that uses it. See the current ways to work together.
Foxton is explicit that he has no interest in turning any of this into a career. He made one editorial with Nick Knight for the first issue of Numero NY and otherwise does it for pleasure, and he says making it a job would take the pleasure out. Fair enough.
But his work looks like something and most brand AI imagery looks like nothing, and that has very little to do with taste. It is input. He spent forty years building a private reference library and is now drawing on it. Your brand has been building one too, by accident, and filing it by season.
The prompt was never the hard part. It was just the visible one.