AI and images

Will AI replace the packshot? Product photography after generators

Sailors in yellow waterproofs stand on a yacht beneath dark clouds and a red ensign.
Crew beneath a red ensign.Photograph by Robert Harper. Robert Harper Photograph Library, Archive.London. From the archive, shown for illustration.

Image generators have taken over the kitchen counter, the marble slab and the sunlit windowsill. They have not taken over the toaster. The packshot survives as the one picture in the campaign that has to be true.

Generators have replaced the set, not the subject: the more scenery is invented, the more the one real photograph of the product has to carry.

When Amazon launched AI image generation for advertisers on 25 October 2023, its example was a toaster. Shown on a kitchen counter beside a croissant instead of on a plain white background, the company said, a product can draw click-through rates 40% higher than ads with standard product images. Notice what the machine made in that story. It made the counter and the croissant. The toaster was already there. Our angle is that this division has held ever since: generators have replaced the set, not the subject, and the plain product photograph, far from disappearing, has become the one image in a campaign that has to be true and the reference from which the rest is built.

This article is general information, not legal advice. Checked on 18 November 2026.

What is a packshot, and why does it exist?

A packshot is the dullest picture in commerce and the most useful: the product, squared up to the camera, evenly lit, on white. Its job is to show exactly what the buyer will receive. No mood, no props, no story. Catalogue pages, marketplace listings and the corner of every advertisement have depended on it for decades.

It was never the only kind of product picture. Above it sat the still life, in which a photographer lit a perfume bottle or a necklace as carefully as a face, and the lifestyle image, in which the product appeared in a room or on a person. The three have always been priced and commissioned differently. The question of what AI replaces has a different answer for each.

What do AI image generators do well for products?

They are very good at context, and fast. Amazon's tool, as the company described it, lets an advertiser select a product and receive a set of lifestyle and brand-themed images in seconds, refine them with short text prompts and produce several versions to test against each other. Amazon framed it as a remedy for a specific problem: in a survey it ran in March 2023, nearly 75% of advertisers who had struggled to build successful campaigns named creative development and choice of format as their biggest hurdles.

That is a real gain. A location, a set build, props, a stylist and a day's shooting are replaced by a prompt. For a small seller who could never have afforded a kitchen, it is the difference between having a lifestyle image and not having one. Generators are similarly useful for seasonal variations, different crops for different placements, and testing which setting sells.

Task Image generator Studio photograph
Background or lifestyle scene Fast, cheap, endlessly variable Slow and costly; real light and real materials
Variations for testing Produced in seconds Each one is a new set-up
The exact product: label, proportions, colour Plausible, not guaranteed identical Records the object as it is
Reflective and transparent materials Convincing in general, unreliable in particulars Controlled and repeatable
Evidence of what was sold None; no object stood in front of a lens The photograph is a record
Legal responsibility for accuracy Remains with the advertiser Remains with the advertiser
Copyright in the result Uncertain or absent, depending on country An author and a chain of title

What do generators get wrong?

The specific. A text-to-image model has learned what perfume bottles look like in general and will produce a handsome one. A brand does not sell perfume bottles in general. It sells one, with a particular shoulder, a particular cap, lettering in a particular typeface, and liquid of a particular colour. A generated diamond ring will sparkle; whether it has the right number of claws holding the right cut of stone is another matter.

This is why the commercial tools are built the way they are. They begin from the advertiser's own product, drawn from its listing or from a photograph, and generate what surrounds it. The hard materials make the point most clearly. Glass refracts whatever is behind it, polished metal reflects the whole room, and a faceted stone does both at once. A studio photographer controls those effects deliberately, as our pieces on lighting jewellery on film and photographing fragrance bottles describe. A generator approximates them, and when the surroundings are invented the reflections in the real product no longer match the scene it has been placed in. A viewer may not be able to say why such a picture looks wrong, only that it does.

What do advertising rules say about AI product images?

In the UK, that nothing changes. The Advertising Standards Authority and the Committee of Advertising Practice set out their position on 31 August 2023: if an ad falls within the ASA's scope, "our rules will apply, regardless of how the ad was created". The regulator judges how consumers are likely to understand an ad, not the tools used to make it.

Three points from that guidance bear on product images:

  • Accuracy. Where AI images are used to make claims about what a product does, they could mislead if they "do not accurately reflect the efficacy of the product". The ASA put this in the same category as photoshopped images and social media filters, a line it has policed for years in cosmetics advertising.
  • Responsibility. Citing an earlier ruling, the ASA said that automating a campaign does not move the duty: advertisers "still have the primary responsibility to ensure that their ads are compliant".
  • Disclosure. The guidance does not set out a duty to say that AI was used, though it notes that how an ad was made might be relevant in some cases.

At the date of that guidance the ASA said it had not, to its knowledge, ruled on an ad using AI-generated images; we have not been able to confirm the position since.

The EU has gone further on labelling. Article 50 of the AI Act, which applies from 2 August 2026, requires providers of generative systems to ensure that outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. An amending regulation published in summer 2026 gave providers of systems already on the market before that date until 2 December 2026 to comply with the marking duty, as summarised by Orrick on 29 July 2026. The exception is telling: the duty does not apply where a system performs an assistive function for standard editing or does not substantially alter the input. Cleaning up a real packshot sits on one side of that line. Generating a new scene sits on the other. The metadata vocabulary already draws the same distinction: the IPTC's Digital Source Type list has separate terms for a digital capture, for media created by a trained algorithm, and for a composite that includes algorithmic media, which is what a photographed product in a generated room is. Our comparison of SynthID, C2PA and visible marks covers how such labels travel with a file.

There is a commercial wrinkle as well. A purely generated image may have no copyright in the United States and an uncertain owner in the UK, as we explain in who owns an AI-generated image. A brand that wants an image nobody else can use still has a reason to commission a photograph.

So will AI replace product photography?

It is replacing the middle of the market and leaving the two ends.

The middle is the routine lifestyle image: product in a pleasant setting, of no lasting value, needed in quantity. Generators do this well enough, and the Amazon example shows why platforms want sellers to have it.

One end is the accurate record. Somebody has to establish what the product looks like, and for a physical object that still means putting it in front of a camera, or building a faithful 3D model from measurements and photographs. Every generated variant descends from that record. As invented scenery multiplies, the record matters more, because it is the only part of the image a regulator, a customer or a returns department can hold the seller to.

The other end is the picture made with intent. The still life that gives a fragrance its character, the jewellery image a house uses for a decade, the cover: these are acts of authorship. They are commissioned for judgement, and they produce an image with an owner.

That is our reading of the published evidence, not a forecast, and it could change as tools improve.

Why are real product photographs still the reference?

Because models learn from them and are measured against them. A generator knows what brushed steel or cut crystal looks like only because photographers lit brushed steel and cut crystal. Research data sets make the dependency explicit. Amazon Berkeley Objects pairs 398,212 catalogue images of 147,702 products with attributes such as material, colour and dimensions, and exists so that models can be trained and tested on real merchandise; our article on product photography data sets for multimodal models goes into the detail.

Studio still life is unusually good reference material. The subject is isolated, the lighting is controlled and recorded in the image itself, and the difficult materials are handled on purpose. Period matters too. A photograph of a 1990s fragrance bottle or mobile telephone shows an object as it was, which no generator can verify.

What the archive holds

The Robert Harper Photograph Library is, in large part, a product photography archive made the slow way. Robert Harper assisted the still-life photographer Lester Bookbinder at the Condé Nast studios, and his commercial clients included De Beers, Harrods, Selfridges, Boots, Heinz, Diageo and Montblanc. The inventory lists 20 folders in its jewellery group (1,607 shots) and 26 in fragrance (1,028 shots), the latter including one folder described as squared-up, full-frame shots and another of group shots, alongside 43 folders of shoes and 35 of cosmetics. The principal format is 6 x 7 colour transparency film, and more than half of the archive is still on film and has yet to be scanned.

Two cautions. Brands and houses named in the catalogue are product subjects; no endorsement or clearance is implied, and branded goods raise trade mark and design questions that have to be reviewed for any use. And for anyone preparing product scans for the web, simple edits remain simple: the background colour remover knocks out a plain backdrop in the browser, with no generation involved.

AI and data teams looking for real studio product photographs as training or evaluation material can read how licensing works on the AI training data licensing page; terms are negotiated directly with the single rights holder, subject to rights review.

Frequently asked questions

Will AI replace product photographers?

It is replacing parts of the job. Generators are widely used for backgrounds, lifestyle scenes and variations, work that once needed sets and locations. The accurate image of the product itself is harder to replace, because advertising rules require it to reflect what is sold and because generation tools typically start from a real product image or product data.

What is a packshot?

A packshot is a plain, accurate photograph of a product or its packaging, usually on a white or neutral background and squared up to the camera, made to show exactly what the buyer will receive. It is the standard image on product pages and the reference from which advertising, catalogue and marketplace pictures are derived.

We know of no specific UK ban, and the ASA's guidance of 31 August 2023 does not set out a duty to disclose AI use. The guidance says the ordinary rules apply however the ad was made: it must not mislead, and the advertiser remains responsible. An image that shows a product as different from, or better than, the real thing risks breaching those rules, and other consumer protection law may apply too. This is general information, not legal advice, and advertisers should take their own.

What do AI image generators get wrong with products?

The difficulties are with exactness. A generator produces a plausible object, not necessarily the specific one: lettering, logos, proportions, the number of facets on a stone or the precise colour of a liquid may drift. Reflective and transparent materials are demanding. For that reason commercial tools usually keep the photographed product and generate only its surroundings.

Why do AI developers want real product photographs?

Because they are the ground truth. Models learn what materials, shapes and packaging look like from photographs of real objects, and they are evaluated against them. Research sets such as Amazon Berkeley Objects pair hundreds of thousands of catalogue images with product attributes. Controlled studio lighting on isolated objects is especially useful for reflective and transparent materials.

Sources

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