AI training data and image licensing

Why product and still-life photography matters to multimodal models

A Chanel No. 19 spray bottle with a gold top and bubbles in the liquid.
Chanel No. 19 voile parfumé.Photograph by Robert Harper. Robert Harper Photograph Library, Archive.London. From the archive, shown for illustration.

A diamond on black velvet is a difficult picture to take and a difficult one for a model to learn. Studio still life is where material, light and description meet.

A well-cut colourless diamond shows the camera very little of itself. What the lens records is the room: every light, reflector and black card around the stone, folded through its facets. Photographing it well is a problem of controlling reflections, and it is one of the reasons studio still life has always been a specialist's trade. It is also why this kind of picture interests people who build multimodal models. The things that make it hard to shoot are the things a model has to learn.

This article looks at what product and still-life photography contributes to vision-language and image-generation systems, what researchers already use, and where a studio photographer's archive fits. It makes no claim about model performance; fit is something a buyer tests on a sample.

This article is general information, not legal advice. Checked on 17 November 2026.

What do multimodal models learn from a product photograph?

Modern vision-language models learn from pairs. OpenAI's CLIP paper, posted in February 2021, described training on 400 million image-text pairs with a simple task: predict which caption goes with which image. Everything the model comes to know about what "brushed steel", "faceted" or "patent leather" look like, it learns from pictures that arrive with those words attached.

That makes the relationship between picture and description the thing that matters. A street photograph captioned "London, 1976" contains hundreds of objects the caption never mentions. A studio photograph of a single watch on a plain ground, described as such, has almost no gap between what is shown and what is said. For teaching attributes, that tightness is useful.

Why is studio still life a distinct kind of data?

Four properties set it apart from general photography.

Isolation. One subject, or a deliberately arranged few, against a controlled background. Segmentation is easy and the subject is unambiguous.

Controlled light. The photographer has decided where every highlight and shadow falls. Lighting is consistent within a job and legible across jobs, which is what a model needs if it is to separate an object's shape from the light on it.

Difficult materials, concentrated. Polished metal, cut glass, liquids, gemstones, lacquer and fur are the daily subject matter of jewellery, fragrance, watch and drinks photography. Reflective and transparent surfaces are well-known hard cases in computer vision, and studio work gathers them in one place.

Legible detail. Product photography exists to show engraving, stitching, typography on a label, the texture of a surface. It is typically shot on larger formats and reproduced large.

Research datasets show the appetite for this kind of material. Amazon Berkeley Objects, released by researchers at Amazon and UC Berkeley, contains 147,702 product listings with 398,212 catalogue images, each listing carrying structured attributes including brand, colour, material, dimensions and style. Its authors used it to benchmark tasks including material estimation, which is the still-life photographer's problem restated for a machine.

Who uses product photography datasets?

The commercial pull comes from applications where goods have to be recognised, described or rendered accurately: visual search, automated product description, retail and advertising imagery, virtual try-on, and quality control of generated images. Developers serving those markets increasingly state where their data comes from. Bria, for instance, says its training assets come from commercial agreements with more than thirty data partners. Shutterstock's data-licensing page invites buyers to discuss their data requirements with its AI data team.

Subject matter is one of the stated drivers of value. Paul Melcher, writing for Kaptur on 12 June 2025, listed resolution, subject matter and exclusivity as the factors that move image dataset pricing. A coherent body of high-end studio work is a subject-matter proposition, not a volume one, which is the argument of our piece on where a 95,000-image archive fits between fine-tuning and foundation training.

What does an editorial and advertising archive add to catalogue images?

E-commerce catalogue photography is abundant and uniform: white background, flat light, standard angles. That uniformity is useful, and it is also a limit. Photography commissioned for magazines and advertising treats the same kinds of object differently. The lighting is expressive, the composition is designed for a page, and the object is often shown with props, surfaces and settings that say something about how it was meant to be seen.

It also carries time. A catalogue image describes a product on sale now. A photographer's archive records what a perfume bottle, a telephone or a pair of shoes looked like in the year it was shot, in the visual language of that year. For a model expected to place objects in a period, or simply to know that products change, dated originals are evidence that current stock cannot supply.

What is in the Robert Harper Photograph Library?

Robert Harper spent two years at the Condé Nast studios assisting David Bailey, Helmut Newton and Lester Bookbinder before setting up his own studio, and was a regular contributor to Elle and The Telegraph Magazine for more than fifteen years. His commercial clients included De Beers, Harrods, Selfridges, Boots, Heinz, Diageo and Montblanc. The archive that resulted includes many of the studio subjects discussed here.

Of its 24 principal categories, the following are still life or product:

Category Examples noted in the inventory
Jewellery De Beers diamonds and the Millennium Star diamond
Fragrance Bottles from Chanel, Calvin Klein and other houses, as product subjects
Watches Catalogued as a principal category
Designer footwear; Handbags and bags Designer shoes, couture accessories
Food; Drinks; Kitchen and tableware Catalogued as principal categories
Corporate and technology Early mobile telephones, microchip development, credit cards, Walkmans and hi-fi
Toys and childhood Toys

Two facts about the physical material matter to a data buyer. The analogue holdings, 52,684 images by inventory count, are mostly medium-format colour transparencies, a format with a much larger image area than 35mm. And those holdings largely await digitisation, so today's deliverable is the 42,803 digital images, with film to follow as it is scanned. What a given film format yields in pixels depends on the scan; the film scan resolution calculator gives the arithmetic. Per-category counts are not published because collections overlap and the inventory has not been checked frame by frame against the physical material.

The rights caveat that comes with branded goods

Product photography has a rights question built into it. The copyright in the photograph belongs to its owner. The name on the bottle, the logo on the dial and in some cases the shape of the object itself may be protected separately, as trade marks or registered designs belonging to the brand. A licence of the photograph does not carry those rights with it, and a brand appearing as a subject implies no endorsement by that brand.

That is manageable, but it has to be managed. The usual options are to flag branded images so the buyer can decide, to exclude them from a licensed set, or to confine a licence to categories where the question does not arise, such as food or unbranded objects. Our article on what a rights review covers before images are licensed for AI explains how those decisions are reached. Archive.London makes no clearance claim for any product image; licences are negotiated directly with the single rights holder, subject to rights review.

Testing the fit

The honest way to find out whether thirty-five years of one photographer's still life is useful to a model is to try some. For ML researchers and data leads working on product, retail or material understanding, Archive.London can supply a dataset card and a sample set under NDA drawn from already-digital, rights-reviewed images, and can discuss first-look arrangements on studio categories as the film is scanned.

To see the documentation and ask for a sample, visit the AI training data licensing page.

Frequently asked questions

What is a product photography dataset?

It is a collection of photographs of goods, usually shot in controlled conditions, paired with text or attributes describing each item. Research examples include Amazon Berkeley Objects, with 398,212 catalogue images. Developers use such sets to teach or test models on object appearance, materials and attributes for uses such as visual search, product description, retail imagery and image generation.

Why are studio product photos useful for training multimodal models?

Because the subject is isolated, evenly lit and deliberately composed, the link between the picture and its description is unusually clean. Studio work also concentrates difficult materials, such as polished metal, cut glass and gemstones, which models must handle for retail and luxury applications. Whether a particular collection improves a particular model is an empirical question to be tested on a sample.

Can photographs of branded products be licensed for AI training?

The photographer's copyright can be licensed by its owner, but brand names, logos, packaging and product shapes may be protected separately as trade marks or designs belonging to the brand. A licence of the photograph does not include those rights. Suppliers therefore review branded images case by case, flag them, or exclude them, and buyers should ask which approach has been taken.

Does film-originated still life differ from digital product photography?

The capture medium differs: a scanned medium-format transparency records colour and tone through film and a scanner, where a digital file comes from a sensor. Some researchers want film-originated images as a distinct, dated sample. No general quality claim follows from the medium; scan resolution, colour management and condition decide what the files are worth, and a buyer should judge samples.

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