What It Takes to Put AI Imagery into Production

Consider a hypothetical commission. A retailer receives front photographs of a coat and a blouse from one of its suppliers. The launch requires product-page images and a campaign, including a reverse view of the coat and an on-model depiction of the blouse. The coat's back is not visible in the supplied photograph. No other reference has yet been supplied.

The request looks like a small addition to an ordinary production brief. It contains an unanswered question about the coat. Someone could generate a plausible back vent, but the front photograph would provide no basis for confirming whether the coat has one or how it is cut. An additional photograph, an approved specification or a sample might settle the question. Until then, part of the commission remains unsupported.

The blouse presents a different task. Its front photograph may support a depiction on a model, provided the chosen method can preserve the relevant garment details. That would still need to be checked and would not verify physical fit. Some campaign adaptations may also be made from the views already available. The retailer therefore has several decisions to make within what initially looked like one job.

In Visual Abundance and Brand Authority, I examined how cheaper image generation changes creative work and the responsibility of those directing it.¹ This essay follows a more specific commission through production. It asks how missing references are resolved, what a supplier owes when work fails, and how much of the required delivery is actually completed. Those questions determine what a brand or retailer can reasonably put into recurring use.

Photoroom's account of its work with Decathlon offers a concrete example of the narrower task. The case study, dated 27 February 2026, describes standardising existing product images through presets and an API connection to the retailer's digital asset management system, using a Google Sheets interface. Backgrounds, product positioning and shadows were among the concerns. It documents an editing workflow around material already held by the retailer. It does not establish the reliability of generating new views of clothing or placing an entire assortment on synthetic models.²

That distinction changes the commission. In an editing task, the supplied photograph can provide a substantial basis for comparison. In a new depiction, the reviewer may need information that the photograph never contained. A format adaptation can preserve the depiction while changing how much of it the customer sees.

These are working distinctions, not a ranking of difficulty. An intricate cut-out may demand considerable expertise. A new scene may be straightforward when the required product information is available and the chosen workflow performs well.

FASHN's documentation makes a related functional distinction. Its Product to Model route can create a new depiction using a face reference, while its comparison with Try-On concerns placing a product into an existing photograph. These descriptions do not promise identical behaviour across functions or prove that every product pixel remains unchanged.³ A merchant commissioning imagery also needs a different assessment from a shopper trying to visualise an outfit. Neither a convincing presentation nor a product name establishes verified physical fit.

For a real launch, the required set might include several views per product, a feed image and a campaign composition. The same master may contribute to more than one delivery. It is still necessary to know whether each required use has been covered. Software that combines the operations may reduce handling, but its combined interface does not make their evidence interchangeable.

This is also where technical feasibility belongs. Before asking who approves an image, establish whether the proposed route can produce it to the required standard. If significant errors cannot be avoided or detected dependably, a better brief or more generous remedy cannot make publication sensible. The appropriate next step may be a narrower task, further testing or conventional photography.

The useful question is who can resolve the missing fact before the delivery deadline. Sending the same front photograph through more generations cannot establish which of several plausible rear constructions belongs to the actual coat. A detailed specification may establish the construction, yet leave uncertainty about its visual appearance. The adequacy of the reference depends on what the requested view is meant to show.

A brand producing its own collection may have direct access to the people who specified the coat's construction. A retailer carrying that brand may need an answer from the supplier. Those are possible routes, rather than fixed advantages attached to company types. A retailer with a well-maintained image feed and clear supplier permissions may be better equipped for a particular task than a brand whose references are scattered across agencies and product teams.

It helps to separate knowledge from permission and from payment. The person who knows the product may not control the photograph. The party commissioning another view may not be the party that gains from using it. For a retailer, an apparently simple reference request can therefore contain a commercial decision about who supplies additional material and who absorbs the work. There is no reliable average in the reviewed evidence for that effort or for supplier response times. Both belong in the assessment of the actual assortment.

The blouse raises related questions about what a reference can establish. A generated presentation on a body may suggest drape, but it does not by itself verify how a particular size fits a customer. If the blouse is made from a sheer or lustrous fabric, the team also needs to distinguish changes caused by lighting and what is worn underneath from changes to the material being offered. These are matters to examine in a trial, not claims that every garment with those properties will fail.

Claid's fashion API guidance illustrates why input requirements deserve attention before work is submitted. It recommends suitable flatlay or ghost mannequin photographs and describes on-model inputs as less predictable. It warns that transparency, strong reflections or substantial occlusion may require better source material and additional iterations. Its linked documentation also describes completing missing outfit elements. That is a generation capability, not evidence that an unrecorded feature of the saleable product has been recovered.⁴

Once a reference gap is identified, the production plan should give it a disposition. Obtaining more information may allow the original task to proceed. A view that makes no unsupported claim may offer a useful alternative. Conventional capture may remain the most economical way to supply the missing image.

Otherwise the production system can appear successful precisely because its unresolved work has disappeared from view.

There is a practical difference between an item rejected before processing and one that fails after several attempts. The first reveals an entry condition. The second reveals a problem encountered within the proposed route. Both can leave a hole in the launch, but they invite different action. Better intake may address the first. Further intake work will not necessarily fix a model that misrepresents a well-documented product.

New photography is not automatically the answer to every gap. It also requires preparation and checking, and an unsuitable photograph can mislead. What matters is having evidence adequate for the particular representation, then testing whether the chosen method can reproduce it. The size of a reference folder is a poor substitute for that judgement. A small, authoritative set may settle the relevant questions more effectively than many attractive images taken for another purpose.

Photoroom's public enterprise offer begins with tests on customer samples. It describes agreed product-fidelity criteria, checks against those criteria and regeneration or credit for qualifying misses. Aesthetic choices and inputs the service cannot process fall outside the stated coverage. Its FAQ links payment to passing the agreed criteria. The public page does not supply an executed customer schedule or a neutral mechanism for resolving a disputed pass.⁵

Pixelz's AI service addendum, updated on 27 August 2026, describes a different arrangement. It provides for discretionary quality checks without an accuracy warranty. Orders are payable when placed, and completed generation fees are earned on delivery rather than customer approval. The text does not promise that another generation will fix a defect. Its liability provision refers to the fee for the specific generation concerned.⁶

These documents compare an advertised enterprise offer with published AI service terms. They establish neither a performance ranking nor the terms an individual customer will negotiate. The narrower insight is that a customer may purchase different treatment of failure alongside the image-making capability. That treatment belongs in the production decision.

Consider an illustrative disagreement in which the blouse is correctly depicted but the model's pose is unsuitable for the intended campaign. A product-fidelity test could pass while the creative director declines the image. If the agreed service concerns fidelity, that rejection may not trigger the same remedy as an altered product detail. The campaign still needs an image, and someone must decide whether to continue generating, change the direction or use another production route. This is a consequence of the distinction, not a reported dispute involving either supplier.

The customer's own authority also needs enough definition to make acceptance workable. A supplier may be able to confirm the construction of the coat's back without being entitled to choose the retailer's campaign setting. Within a brand, the person responsible for construction may make a different judgement from the person responsible for presentation. The work need not pass through a new committee. It does need an answer when one decision has been made and another remains open.

The most useful commercial discussion therefore starts with a specific failure. If a print on the blouse is misplaced, what reference settles the matter? If an image passes a supplier check but fails the customer's comparison, how is that difference assessed? A deadline adds another dimension. A further attempt can be a satisfactory remedy early in production and of little practical use after the launch. Agreeing what is owed in that situation makes the production promise easier to interpret.

Nor does the existence of a check establish that it catches every material error. A false acceptance can leave an incorrect image in the approved batch. The decision to rely on a checking process needs evidence of what it examines and what can escape it. A supplier's promise to inspect work is relevant, but it is not a measurement of inspection accuracy across the customer's assortment.

The financial consequences extend beyond the generation charge.

The public material reviewed here does not establish reimbursement of all those costs. If a proposed agreement does cover them, that coverage should be read in its own terms, including its exclusions and limits.⁷

Paying for attempts can still be reasonable. A team may value flexibility, have strong internal editing capability and find the complete route economical. A more extensive acceptance commitment may carry eligibility conditions or a price that changes the calculation. Neither arrangement deserves preference before the task, expected work and actual commercial terms are understood.

Experienced production teams already make comparable distinctions when commissioning photographers and retouchers. There is no benefit in giving familiar responsibilities new names unless something material has changed. Here, the changes worth examining are the billable event, the scope of the promised result and the remedy when it is missed. If existing arrangements cover them adequately, the additional work may be small. Acceptance should help people commission the job clearly, without becoming a substitute for doing it well.

That does not make AI processing automatically prohibited. It means the proposed route needs to be compared with the permission already held. An agreement may cover the intended changes and publication. It may require clarification or an extension. Section 23 adds the question of consent for publishing or exploiting an adaptation, subject to its statutory boundary, rather than declaring every technical image adjustment unlawful.⁹

For the hypothetical retailer, asking a supplier to confirm the coat's back vent would resolve a product question. It would not necessarily extend the licence to upload a photograph to another service or publish a transformed version in a new territory. Conversely, broad editing permission would not supply the missing product fact. These questions can travel together in the same request while still requiring different answers.

An identifiable person introduces further considerations. Permission for the proposed likeness use must be assessed separately from permission to use the photograph. German portrait law sets a consent principle with limited exceptions. Where personal data are processed, the GDPR requires attention to the legal basis and the parties' actual roles. A processor arrangement and any international transfer need their own applicable safeguards.¹⁰ Paying for an earlier shoot cannot be treated as conclusive permission for every subsequent synthetic depiction.

Supplier data terms can make the route more specific. FASHN's terms dated 7 August 2026 restrict training and other model improvement with customer content unless the customer expressly opts into a separately agreed service. Processing to deliver the service remains permitted under the stated conditions. Its data-processing agreement and service-specific subprocessor information therefore still matter.¹¹ A no-training commitment answers one question about use of the material. It does not mean that no external processing takes place, or that every service follows an identical geographical route.

Even when no personal data are involved, an unreleased collection may require confidentiality. That concern belongs in the choice of service and agreement. It should not be dismissed because an image contains no face, nor translated into a claim that all product uploads carry the same exposure.

Publication brings a separate EU transparency question. Article 50 distinguishes machine-readable marking by providers from disclosure by deployers where content constitutes a deep fake. The deployer duty has applied since 2 August 2026. Relevant systems placed on the market before that date have a provider marking transition until 2 December 2026. The transition does not postpone deployer disclosure.¹²

The image's content and presentation matter. An invented model does not automatically put an image outside the deep-fake definition, which also concerns existing objects. The assessment includes whether the content would falsely appear authentic or truthful to a person viewing it. Equally, an on-model image is not categorically a deep fake merely because generation was used. The limited provider exception for standard editing cannot become an exemption for all packshots. Human review of certain public-interest text is not a general image exception. Evidently creative works have an adapted form of disclosure, rather than a blanket release from it. Where deployer disclosure is required, embedded metadata alone does not make that disclosure perceptible to the viewer.¹³

These distinctions allow a concrete decision about a particular production and release. They do not produce a universal label that clears the image. Under Germany's unfair competition rules, a pictorial representation can still mislead about a material product characteristic where the statutory conditions are met.

Some uncertainty also remains after contractual permission has been established. Rights assigned by a service do not necessarily create copyright protection, exclusivity or freedom from third-party claims. FASHN's output provisions expressly qualify what is allocated.¹⁵ A brand may therefore have permission to use an image without securing the degree of protection it wants for a distinctive campaign. The consequence could be a different production choice for that campaign, while a less exclusive catalogue use proceeds under the same service.

A crop provides a simple illustration. If a product specialist approved a view of the blouse because its cuffs and hem were visible, a tighter crop may remove the detail that made the view informative. The underlying product depiction might remain accurate. The derivative could still fail the purpose for which that view was required. A layout adjustment that leaves those properties intact presents a different case.

The practical question is what the earlier approval depended on.

A new generated region touching the product needs closer attention. Copy that introduces a material or performance claim requires a basis beyond the image's visual approval. Keeping those dependencies visible allows review to follow the affected decision, rather than restarting the whole process or assuming that all prior approvals remain valid.

Google Merchant Center supplies a public example of why destination matters. Its rules for the main product-image attribute restrict promotional elements and overlays. They also require provenance metadata for images created with generative AI and instruct users to preserve embedded markings such as the IPTC DigitalSourceType property.¹⁶ A campaign composition containing promotional copy may therefore need a separate export for the product feed. These are rules for that submission route, not a general description of every Google advertising format.

The difference is operational. The campaign file can meet its own brief while remaining unsuitable for reuse as a feed image. Taking the headline away may address the overlay issue. It does not establish that all other requirements have been met. Nor does a correct export at one stage prove that embedded provenance survives later processing or delivery. That needs to be checked through the actual route. No such export test was carried out for this essay.

An approved template can make this work more economical. Where resizing and placement are deterministic, and the relevant properties are preserved, the team may be able to reuse established decisions with little additional effort. Generating a new background, extending a product region and moving text within a fixed layout should not be treated as the same kind of change simply because each is available through a button.

For a multi-brand retailer, a new line of copy may make a claim that only the supplier can substantiate. A destination in another market may also change the permissions needed. A brand can face similar issues between a central creative team and local release. The useful record is therefore more than an approval flag. It connects the approved version with the use and information on which the decision rested.

This proposed approach depends on knowing which decisions each transformation preserves. Adobe and Canva offer production and adaptation capabilities, but the reviewed descriptions do not show whether a particular product approval remains valid after an edit.¹⁷

The benefit of that understanding is that it can reduce unnecessary repetition while exposing the edits that matter. An earlier approval becomes reusable evidence for a defined purpose. The cost of reopening it, or preparing another required export, then belongs to the production task. It is part of what has to happen before the campaign and product feed are both supplied.

The example gives a useful account of one bill. It leaves open how many required views can be produced, which inputs need attention and how much customer work is involved. Brandmachine's enterprise structure, which combines a platform fee with usage and no export fee, changes that bill again. Its €3 to €4 indication depends on volume and scope.¹⁹ Neither amount is a complete production cost.

The appendix retains the bounded rates for FASHN, Uwear and Photoroom. Their units differ. Credits buy specified operations, a successful generation can still be creatively unusable, and an export can remain unpublished. Converting those prices into a league table of finished images would hide the very work the comparison needs to examine.

The required set provides a more stable starting point. Fix the products, views, destinations and necessary language versions for the operating period, with the quality threshold and deadline. Then account for what was supplied through AI, through a hybrid route and conventionally. A delivery position should count once according to its final route. Producing several candidates for it does not enlarge the requirement.

A cost per accepted deliverable can be calculated from the complete assignment, including unsuccessful work and replacement production. Here, acceptance means that a required delivery has passed the necessary product, creative and placement checks. It is distinct from a supplier deciding that an output is billable.²⁰

The figure needs a completion account beside it.

Report A/R with the unit cost. If publication by the deadline is part of the job, report P/R as well. Approved work can still be late or unused. Where several views are needed for each product, also show how many products have their complete set. Numerous front images cannot compensate for a missing reverse view when that view remains a launch requirement. Withdrawn or replaced versions should remain traceable over the intended period of use.

A source excluded during feasibility screening stays in the required set until an explicit scope decision changes it. The company may sensibly reduce that set, but it has then changed the assignment. The original requirement and revised requirement should both remain visible. If nothing is accepted, the accepted unit cost is undefined. The expenditure and the outstanding job remain real.

The numerator needs equal care. Each cost must enter once. In this method, a separately commissioned conventional replacement belongs under exceptions, including the work within that replacement. A retry and its repeat review within the continuing route belong under repair. First-pass service fees and first review belong in their corresponding categories. Reviewer hours can be reported across those categories, but their total must not then be added again. A credit reduces the original charge once. A disputed refund is not yet a saving.

Fixed costs should be allocated over a realistic operating period and usable capacity. A platform commitment cannot be made negligible by dividing it over an unlimited stream of hoped-for images. The same discipline applies to setup work and existing photographs. Historical expenditure should not be charged again to one option while treated as free in the other. New payments and the consumption of existing resources need to be distinguished on a consistent basis.

This also separates cash from capacity. Avoiding an external invoice changes spending. Releasing staff time changes available capacity unless a corresponding expense falls. Both can matter, but they support different decisions. Valuing internal hours and adding the same salary cost again would count the resource twice. Faster delivery, conversion and returns require a separate account of commercial effects rather than being relabelled production savings.

The research reviewed for this essay did not establish a comparable independent full-cost study published after 24 July that supplies an industry saving for these tasks.²¹ A company does not need such a study before running a modest trial. It does need to observe the material costs and unresolved work in its own commission. Where the full calculation would cost more than the decision warrants, a proportionate account with explicit uncertainty is enough to guide the next small commitment. A recurring seasonal programme deserves firmer evidence than an afternoon's experiment.

The account supports a limited conclusion about recurring use in an identified range. Its undated publication and undisclosed production ledger leave the timing and full economics unresolved.

For a team considering its next commitment, the relevant question is what would change. Repeating a known task with comparable references can draw on earlier work. Adding another supplier, a different material or a new destination may disturb an assumption on which the result depended. The appropriate examination follows that change. It need not repeat every previous test, but it should not infer coverage from the number of images already produced.

A useful trial consequently retains the products that were excluded and the conditions under which the successful work passed. The team should be able to explain whether a missing view required another source, whether repeated attempts remained economical and whether the final exports could be used.

A smaller business may already have a dependable studio arrangement that covers the required set cheaply. The additional preparation and service management could outweigh the benefit of changing it. Another business may find a useful division of work, with conventional capture providing authoritative detail and AI supporting a defined set of presentations or adaptations. These are possible decisions to evaluate, rather than measured outcomes invented to balance the adoption cases.

A concern about brand image deserves similar specificity. The person raising it may be protecting a feature of the commission that has not yet been stated clearly, such as a requirement to depict an actual person or occasion. It may also be an assumption that a bounded test can examine. The available evidence does not establish how common either motive is among creative leaders. Describing hesitation as resistance would settle the diagnosis before understanding the work.

The reviewed cases likewise provide no comparable regional denominator from which to rank European, American or Asian adoption.²³ A company gains little by treating an unsubstantiated account of national progress as a deadline for its own production decision. Its next commitment can be ambitious and still have a defined scope. Learning comes from knowing what that commitment tests and what result would justify changing it.

What can be established is the basis on which the choice should rest. The required views need adequate references. The commissioned service needs intelligible terms for delivery and failure. Changes after approval need to be assessed against the decisions they affect. The cost account needs to show the job still outstanding alongside the work already accepted.

These conditions do not guarantee a successful campaign. They make a recurring production commitment easier to explain and revise.

I would welcome specific experiences from brands and retailers, including what made regular use workable, what led to a narrower role for AI and when continuing with photography proved the better choice. What changed the decision, and what work remained unresolved, would be particularly useful to understand.

This is a selective map of the providers already examined in the research, organised by production task. An offer describes what a supplier advertises. It does not establish tested suitability for the reader's assortment. Unknown means not established in the reviewed material. App, API and managed-service arrangements remain separate. No ordering implies preference.

Product depictions and specialist fashion workflows

Editing and managed production

Broader adaptation and contextual routes

Provider contracting entities and ultimate ownership were not comprehensively established. FASHN's reviewed terms identify FASHN LTD. Other corporate and order-form details should not be inferred from product names. This appendix records public offers, without certifying general availability of every described feature or the implementation behind it.

Dated charging units

All public prices below were checked on 17 September 2026. No currency conversion or tax treatment is assumed. FASHN's terms exclude applicable taxes unless otherwise stated. No common tax basis has been established across the listed suppliers. A named quality setting is a supplier-specific configuration, not a common quality standard. Calculated examples are arithmetic on the stated fees, without an assumed acceptance rate.

The Brandmachine example can be checked at each supplied setting. It assumes five chargeable renders and exactly one product-image export. Five renders come from the supplier's stated assumption of 1.25 rounds of four images, without a disclosed representative measurement. None of these totals includes all customer production work.

For Brandmachine, technical render-failure billing, repeat-export rules, tax, exact render resolution, measured iteration distribution and a specific enterprise quotation remain unknown.

Claid's monetary package price and the Botika and Veeton plan discrepancies remain unresolved. Standard Pixelz retouch rates are deliberately excluded from the generative comparison. These omissions do not establish higher prices or inferior performance. The table shows the charging events the evidence can support.