BRENNEN

Industries / Retail, hospitality and personal services / Photography

Commercial work

This is photography made to order for businesses: fashion and advertising pictures, product shots for shops and catalogues, food, interiors and corporate work.

5

Milestones achieved, from 1911 to 2025.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1911

    Clothes photographed as a mood

    In April 1911 the French magazine Art et Décoration printed Edward Steichen's photographs of gowns by the Paris designer Paul Poiret. Fashion photographs already existed, and the Fashion Institute of Technology (FIT) holds examples from the magazine La Mode Pratique dating to 1892, but those showed the garment plainly, like a fashion drawing. Steichen's pictures were moody and atmospheric and suggested a way of living around the clothes; FIT's April Calahan writes that many scholars date modern fashion photography from them. In 1922 Condé Nast hired Steichen as chief photographer of Vanity Fair, with regular fashion work for Vogue.

    blog.fitnyc.edusmithsonianmag.com

    Credit

    Edward Steichen; Paul Poiret; Art et Décoration

  2. 2014

    Renders replace product shots

    IKEA's in-house agency began planning a move from photography to computer-generated images about ten years before 2014, to save the cost of building prototype furniture and shipping it around the world to be photographed. Its first computer-generated product image, of the Bertil chair, dates from 2006, and by 2012, 12% of its catalogue and online images were computer-generated, by the Wall Street Journal's count. In 2014 Martin Enthed of IKEA Communications told the trade site CGSociety that 75% of IKEA's single product images were computer-generated, while most room scenes in the catalogue were still photographed in its studio in Älmhult, Sweden.

    LessonSeveral outlets first reported that 75% of the whole catalogue was computer-generated and had to correct it; software replaced the plain product shot first, not the room scene.

    slate.com

    Credit

    IKEA Communications; Martin Enthed

  3. 2017

    A digital supermodel

    Shudu, a computer-generated fashion model designed as a Black woman, was created in 2017 by The Diigitals, an agency of digital models founded by Cameron-James Wilson. NBC News called her "thought to be the world's first digital supermodel" and reported in 2023 that she had booked work with brands such as BMW and Louis Vuitton. Her creator, who is white, drew criticism for her design.

    nbcnews.com

    Credit

    Cameron-James Wilson, The Diigitals

  4. 2023

    AI models draw a backlash

    In late March 2023 Levi Strauss & Co. announced a test of AI-generated clothing models from the design firm Lalaland.ai, to show online shoppers more kinds of people wearing its clothes. Critics said it was a cheap way to look diverse and would take paid work from human models. Levi's denied that it meant to save costs and said the AI models would supplement, not replace, its photo shoots with live models.

    LessonThe plan was judged on who would lose paid work, not on how the pictures looked.

    nbcnews.com

    Credit

    Levi Strauss & Co.; Lalaland.ai

  5. 2025

    Consent for digital models

    From 19 June 2025, New York State's Fashion Workers Act requires a model management company to get a model's clear written approval before creating or using a digital replica of the model. The approval must be separate from the representation contract and state the scope, purpose, rate of pay and length of time the replica will be used, and the company may not create, alter or manipulate a replica without it.

    dol.ny.govnbcnews.com

    Credit

    New York State Legislature; the Model Alliance, founded by Sara Ziff, which campaigned for the law

Watch for

  1. Shop content made by AI

    Zalando, the Berlin online fashion retailer, said in May 2025 that about 70% of its editorial campaign images in the last quarter of 2024 were AI-generated, and that it was piloting digital twins, 3D copies of real models, with Orendt Studios. Presenting its 2025 results in March 2026, co-chief executive Robert Gentz said more than 90% of Zalando's content is now made with AI, giving 70% more content than a year earlier at no higher cost, and that product photography that took about six weeks now takes a few days.

    Achieved whenAn outside audit or study measures what share of Zalando's images are made with AI, what each costs, and whether returns or complaints about how products look have changed.

    fashionunited.comcorporate.zalando.com

    Claimed by

    Zalando, March 2026

Open

  1. 01

    AI garment photos, measured

    A public test of how faithfully AI tools show a real garment when they put it on a generated model or a new background. Measure 100 garments at marked spots with a spectrophotometer, an instrument that reads a colour as numbers. Photograph each garment flat and on a consenting model under controlled light, with a colour chart in the frame. Give the flat photos to several AI on-model tools, then measure the colour error of each result against the garment itself in CIEDE2000 (ΔE00), the standard measure of how different two colours look, and have two people count changed details such as buttons, pockets, seams, prints and logos.

    The brief
    Credit · unclaimed
  2. 02

    Robot loads the packshot table

    A low-cost robot arm, taught from human demonstrations, that takes small, unbreakable products from a tote, sets each one on a photo turntable at the marked spot and facing the set way, and lifts it off after the shots. A packshot rig, an automated setup for catalogue photos on a plain background, would then run without a person loading it.

    The brief
    Credit · unclaimed

Open 01 · the brief

AI garment photos, measured

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public test of how faithfully AI tools show a real garment when they put it on a generated model or a new background. Measure 100 garments at marked spots with a spectrophotometer, an instrument that reads a colour as numbers. Photograph each garment flat and on a consenting model under controlled light, with a colour chart in the frame. Give the flat photos to several AI on-model tools, then measure the colour error of each result against the garment itself in CIEDE2000 (ΔE00), the standard measure of how different two colours look, and have two people count changed details such as buttons, pockets, seams, prints and logos.

Why it is still open

Retailers now publish such images at scale; Zalando says more than 90% of its content is made with AI. Research benchmarks compare the generated picture with a reference photo of the garment, not with the garment. The DAT framework of Kaidong Zhang and colleagues (August 2026) splits garment fidelity into seven dimensions, colour among them, judged against reference images; Hyeryeon Park and colleagues (Fashion and Textiles, June 2026) score garment consistency with vision-language models against the original garment photo; and DiffusionTrend uses CIEDE2000 only to decide when to stop generating. The software company img.ly measured with CIEDE2000 how close 26 image models come to exact colours named in a prompt, which is not the same as keeping a real garment's colour. This research found no test against measured physical garments.

fashionunited.comarxiv.orglink.springer.comarxiv.orgimg.ly

Built with what exists

A handheld spectrophotometer such as an X-Rite Ci64 or a Nix Spectro 2; a Calibrite ColorChecker chart; a camera and continuous lights; open try-on models such as IDM-VTON or CatVTON from Hugging Face, and trial accounts of commercial on-model tools; Python with the open colour-science library to compute ΔE00.

  1. First step

    Measure and photograph ten garments in strong colours, such as reds, teals and navies, and run them through two tools.

  2. Done looks like

    100 garments run through at least three tools, with a published table of median colour error per tool, the share of garments above a stated threshold such as ΔE00 3, the count of changed details per tool, and the same numbers for the studio photographs as a baseline.

Open 02 · the brief

Robot loads the packshot table

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A low-cost robot arm, taught from human demonstrations, that takes small, unbreakable products from a tote, sets each one on a photo turntable at the marked spot and facing the set way, and lifts it off after the shots. A packshot rig, an automated setup for catalogue photos on a plain background, would then run without a person loading it.

Why it is still open

The rigs automate the camera and the turning, not the loading. In PhotoRobot's own account of a shoot, the operator scans a barcode, weighs and measures the product, then places it where a laser cross on the turntable shows. MRMC's page for its Orbital rig says to "mount the item on the turntable" and the rig "will do the rest". This research found no vendor or study that loads the table by robot.

photorobot.commrmoco.com

Built with what exists

The open-source SO-101 arm and Hugging Face's LeRobot software (see the Note "No download"); a policy trained from a few demonstrations, such as LeRobot's ACT or NVIDIA GR00T 2 (see the Note "Five demonstrations per task"); a USB camera over the table; a motorized turntable; ArUco markers read with OpenCV to measure where each item lands.

  1. First step

    Set an SO-101 beside a turntable and record 50 demonstrations of placing ten objects.

  2. Done looks like

    200 placements of at least 40 objects not used in training, with the published share placed within 5 mm and 5 degrees of the mark, items per hour against a person doing the same job, and the number of drops, all on video.