H&M’s July 2025 digital-twin campaign came with something useful: an account of the people involved. The retailer named creative collaborators, published comments from a photographer and model, and released a behind-the-scenes film alongside the images.
That makes the project more interesting than another gallery of convincing artificial faces. There was a production around it. People had made choices about how the images would look and how they would be presented.
In its announcement of the digital-twin project, H&M described the work as an exploration of generative AI. It did not publish figures establishing lower production costs or higher sales. Those questions remain separate from the fact that the campaign appeared.
For other businesses considering synthetic people, that distinction is a sensible place to start. A published image demonstrates an output. Deciding whether the process suits another business takes more work.
Someone still has to approve what customers will see.
The jacket deserves as much attention as the face
Discussions about generated people tend to concentrate on anatomy. Does the hand look plausible? Do the eyes match? Is the expression convincing?
A retailer has additional concerns. The person might look perfectly believable while wearing a version of a garment that the company does not sell.
Take a hypothetical jacket campaign. The generated image introduces an extra seam, changes the cuffs or makes the material appear heavier. None of those differences necessarily makes the picture look strange. A customer could easily read them as product details.
That is precisely why they need checking.
The brief should establish what may change. A background might be open to interpretation. The jacket’s fastening, colour and construction might need to remain faithful to the stock item.
The intended placement matters, too. An imaginative image introducing a collection has a different job from a product photograph used to judge fit. A team should decide which job it is commissioning before selecting its favourite result.
There is room for artificial imagery in early creative work. It can give colleagues something concrete to discuss: a composition, a setting, a direction for the campaign. At that stage, an image can be useful without being ready for customers.
Trouble starts when that distinction disappears. A concept file travels into a presentation, someone likes it, and the next person assumes it has already been checked.
Putting a review status beside the file would be a small improvement with an obvious purpose. The designer receiving it should know whether it is an idea, a working draft or an approved asset.
A fictional face can acquire a factual claim
The identity question goes beyond appearance. An audience also reads the words surrounding an image.
A fictional character might accompany an article or appear in a clearly presented campaign. Add a customer review underneath, however, and the publication starts to suggest that someone actually bought and used the product.
The face may be invented. The claim about the experience still needs a factual basis.
This is worth examining at the page level. A caption can change the meaning of an image that looks straightforward in isolation. So can a profile biography, a location tag or a headline describing an event.
The same applies when a real person’s digital representation appears in a new setting. Viewers might reasonably assume the person attended a shoot or approved the particular appearance. The production team needs to understand which assumptions its presentation invites.
“AI-generated” is useful information, but it leaves several possibilities open. Was the whole person invented? Was the photograph altered? Does the image represent someone who participated in the project?
The explanation can be brief. What matters is whether it answers the relevant question.
Adult-oriented services such as JOI (AI body generator) belong to this wider discussion of synthetic characters. In that setting, identifying fictional material and documenting any use of a real person’s likeness are especially important distinctions. A realistic result does not establish permission.
These considerations also belong in the original brief. Waiting until publication to ask whose identity is involved makes an otherwise manageable decision harder to resolve.
For collaborators, the questions are concrete: which project, which channels, what duration and what kinds of alteration? Applicable legal requirements depend on the circumstances and jurisdiction.
Faster generation can mean a different kind of workload
A generation time is easy to demonstrate. Total production time is harder to show.
It includes the unsuccessful attempts, the review, the corrections and the preparation of final versions. Those stages can involve different people, which makes them easy to overlook when evaluating a tool.
Suppose a small team needs images for a website banner, an email and several social posts. One composition may work well in a wide format but become awkward in a narrow crop. Text needs space. The product needs to remain visible.
Someone has to make those adjustments.
A useful comparison would follow one assignment from brief to delivery. The team could then compare the AI-assisted process with its usual method, using the same definition of “finished”.
There is no need to count every click. A few records would tell a more useful story:
| Record | Why it belongs in the calculation |
| Attempts before selection | Captures work behind the chosen image |
| Editing time | Shows what remained after generation |
| Review time | Includes product and identity checks |
| Finished versions | Establishes what was actually delivered |
| Staff and software costs | Gives the assignment a total cost |
This also makes a subscription easier to assess. The relevant question is how much usable work it supports.
TechStory’s examination of cloud gaming subscription costs considers value in relation to actual use. Creative software deserves a similar calculation: which features does the team need, and how often?
An occasional experiment and a regular production requirement justify different spending decisions. A large allowance of generations may be irrelevant if approval takes most of the available time.
A pilot can expose that bottleneck before the business builds a workflow around the tool.
Keep the explanation with the image
Creative files are handed around. Designers leave projects, agencies change and old assets get reused. The person publishing an image months later may know very little about its origin.
That makes the accompanying record important.
It should identify source material, relevant permissions, the intended purpose and the latest review decision. Where a real person is represented, the record should make the agreed scope of use accessible to whoever handles the next request.
It should also distinguish permissions from product approval. An authorised likeness can still appear beside an inaccurate garment. A visually correct product image can still carry an unsupported caption.
These checks cover different questions, so one approval should not quietly stand in for all of them.
The final review belongs in the actual layout. On a mobile page, a disclosure might become difficult to find. A crop could remove context. A new headline might suggest something the image was never intended to prove.
Before publication, the responsible editor should see the finished combination: image, copy, placement and explanation. That is where the audience encounters the work, and where the team can judge whether it says what they intended.
