Labelling AI-Generated Product Images

The question brands actually face

Most brands asking about AI labelling are not asking a legal question. They are asking a commercial one: if I disclose, does it cost me sales, and if I do not, what am I exposed to?

The honest answer is that it depends on which image you are talking about. A generated on-model photograph of a real garment you actually sell is a very different proposition from a generated image of a product that does not exist. Both may technically be AI-generated. Only one of them misleads a buyer about what arrives in the box.

That distinction is the useful one to organise your policy around, because it is also the distinction most platform rules and consumer protection regimes care about. The concern is misrepresentation, not the tool.

Two different mechanisms, often confused

There are two entirely separate ways an image can be marked, and conflating them causes a lot of wasted effort.

A visible label is pixels. Text composited into the image itself, so anyone looking at the picture can see it. It survives being screenshotted, re-uploaded, cropped by a platform, or pulled into a marketplace listing. It also permanently alters the image, which is why it matters whether it is baked in or applied at export.

Embedded provenance is metadata, not pixels. Content credentials and similar standards attach a signed record of how a file was made. It is invisible, it does not touch the image, and platforms can read it automatically. Its weakness is fragility: re-encoding, resizing and stripping on upload routinely destroy metadata, and many platforms remove it as a matter of course.

Neither replaces the other. Metadata is the machine-readable answer and is easily lost. A visible label is the human-readable answer and cannot be lost, because it is the picture.

Expect the answer to differ per channel

There is no single setting that satisfies everywhere you sell, and treating it as one policy is where brands come unstuck. The variables that actually change the answer:

• Your own site is where you have the most latitude and the most reason to be straightforward, because a returns dispute about "it did not look like this" is expensive.
• Marketplaces set their own image rules and enforce them mechanically. Their listing specs are the document to read, and they change.
• Paid social and search have advertising policies separate from the platform's organic rules, and disclosure requirements for synthetic media in ads have been tightening.
• Wholesale and retail partners often impose contractual image requirements stricter than any law, and those are negotiated rather than published.
• Jurisdiction matters independently of platform. Transparency obligations for synthetic content have been moving forward in several markets, so where your customer is can change the answer.

The practical consequence: you want the ability to produce a labelled and an unlabelled version of the same image without regenerating it, because different destinations will want different things from one asset.

Why the label should be a per-export choice

A watermark burned into every file at generation time seems like the safe default. In practice it creates problems that are worse than the one it solves.

Once a label is baked into the master, it is in every downstream use. It is in the file your wholesale partner puts on their own site. It is in the asset your agency crops for paid. It is in the image a retoucher receives as a starting point. Removing it means either regenerating from scratch or retouching it out, and retouching out your own disclosure is a worse position than never having applied it.

The workflow that holds up:

1. Keep masters clean. The generated file is raw material. Store it unlabelled.
2. Apply the label at export, per destination, on a copy. Toggle it on for the channels that want it, off for the ones that do not.
3. Record the decision, not just the file. Which SKUs were generated, and which channels got labelled exports. This is the part that answers a question six months later.
4. Re-export rather than un-label. If a channel changes its rules, produce a new export from the clean master.

In Jinko Generator the label is a toggle next to the download control for exactly this reason. It composites onto a copy at download time, scaled to the image so it reads the same on a 1K preview and a 4K export, and it never touches the stored result. Toggling it off and downloading again gives you a clean file with no regeneration and no extra cost.

Does disclosure cost you conversions?

This is the fear, and it deserves a straight answer: nobody has published a figure you should rely on, and any specific percentage you see quoted is almost certainly extrapolated from a study that did not measure your category.

What can be said with more confidence is where the risk actually sits. Buyers react badly to feeling deceived, and they react to that far more strongly than to the knowledge that an image was computer-generated. A generated on-model shot of a garment that arrives looking exactly like the picture is not the thing that generates complaints. A picture that flatters a product into something it is not, generated or photographed, is.

The category matters too. Colour and fit accuracy carry most of the return risk in apparel, and neither is affected by whether the model was real. If your generated imagery is faithful to the garment, disclosure is a much smaller commercial question than it feels like.

A default policy worth starting from

Adapt to your own risk posture, but this is a defensible starting position:

• Never misrepresent the product. Colour, texture, construction, hardware and proportion match the physical garment. This is the rule that matters most and it is not a disclosure question.
• Do not imply a real event that did not happen. A generated model in a generated location is a product presentation. An image that reads as documentation of a real occasion is a different claim.
• Do not generate a person who could be mistaken for a specific identifiable individual, including a real model you have previously worked with, without their agreement.
• Label where the channel asks, and where a reasonable buyer would want to know. Have the mechanism ready so this is a decision rather than a project.
• Keep the paper trail. Which images were generated, from what source, for which channel.
• Re-check the rules on a schedule. This area is moving. A policy set once and never revisited is the actual risk.

Applying a label in practice

Whatever tool you use, a label is only useful if it survives contact with real distribution:

• Scale it to the image, not to a fixed pixel size, so it stays legible at 1K and proportionate at 4K.
• Put it somewhere crops do not remove it. Platforms crop to their own aspect ratios, and a label in a corner that gets trimmed is worse than none.
• Keep it readable against any background, which usually means a semi-opaque plate behind the text rather than text alone.
• Mind the format. A transparent cut-out has to stay PNG; re-encoding it as JPEG flattens the alpha onto black, which you discover after a client has placed the file.
• Do not rely on metadata alone if the disclosure actually matters, because uploads routinely strip it.

Frequently asked questions

Do I have to label AI-generated product images?

It depends on where you sell, which channel you are publishing to, and what the image depicts, and the rules are actively changing. This is not legal advice: check the current requirements for each market and platform you use, and take proper advice. What you can do regardless is make sure the mechanism is ready, so labelling becomes a per-export decision rather than a project.

What is the difference between a visible AI label and content credentials?

A visible label is pixels composited into the image, so a person can see it and it survives screenshots, re-uploads and cropping. Content credentials and similar provenance standards are signed metadata: invisible, machine-readable, and easily destroyed because re-encoding, resizing and platform uploads routinely strip metadata. They answer different questions and neither replaces the other.

Should the AI label be permanently applied to my image files?

Generally no. Once a label is baked into the master it appears in every downstream use, including wholesale partners, agency crops and retouching handoffs, and removing it means regenerating or retouching out your own disclosure. Keep masters clean and apply the label to a copy at export, per destination.

Does disclosing that an image is AI-generated reduce conversions?

There is no reliable published figure to depend on, and quoted percentages are usually extrapolated from studies in other categories. The clearer risk is misrepresentation rather than disclosure: buyers react far more strongly to feeling deceived than to knowing an image was generated. In apparel, colour and fit accuracy drive most return risk, and neither depends on whether the model was real.

How does the AI label work in Jinko Generator?

It is a toggle beside the download control. The label composites onto a copy at download time and is scaled to the image width, so it reads the same on a 1K preview and a 4K export, and the stored result is never modified. Turning it off and downloading again produces a clean file with no regeneration and no extra credit cost.