AI product images: generate the scene, never the product

Generative imagery solved a real problem for small catalogues: a shop with four hundred products cannot afford four hundred lifestyle shoots. What it did not solve is the requirement that an advertisement accurately represent what arrives in the box. That requirement predates AI by about a century, and it has not moved.

The workable line is simple, and it holds up both commercially and legally: generate the environment, preserve the product.

What that means in practice

• Fine: replacing a white studio background with a kitchen counter, a beach, a desk. Changing light and shadow. Adding props that are clearly props and are not sold with the item.
• Fine: generating a model wearing a garment, if the garment's cut, colour and fabric are preserved and the image is disclosed as illustrative where required.
• Not fine: regenerating the label, the logo, the text on packaging, the weave of a fabric, the finish of a metal, or the number of items included.
• Not fine: a generated image that implies a capability the product does not have, or a size relationship that is wrong.

The failure mode is boringly practical. Generated labels produce plausible nonsense text. Generated fabric invents a texture. Customers notice, and the cost lands as returns and support tickets long before anyone raises a legal point.

The disclosure rules caught up in August 2026

The EU AI Act's transparency obligations under Article 50 became applicable on 2 August 2026. For marketing teams, two duties matter: providers of generative systems must mark synthetic output in a machine-readable way, and deployers who publish artificially generated or manipulated image, audio or video content that could reasonably be mistaken for real must disclose that it is artificially generated.

There is an exemption where the content is clearly artistic or fictional, and the disclosure then only has to be made in a way that does not spoil the work. A product photograph on a shop page is not that case. If a shopper could reasonably think the picture is a photograph of the actual item in an actual place, the honest reading is that you disclose.

Advertising law did not change, and it is stricter

Separately from the AI Act, consumer protection law in most markets prohibits misleading commercial practices. In the EU this is the Unfair Commercial Practices Directive; in the US the FTC applies the same principle through its deception standard and its rules on endorsements. None of these care how an image was produced. They care whether it misleads.

That is why disclosure is not a licence. Labelling a picture as AI-generated does not make an inaccurate depiction lawful — it makes an accurate one transparent.

Briefing a model so the product survives

1. Start from a real photograph. Image-to-image with a strong structural constraint, not text-to-image from scratch. The source photo is the ground truth.
2. Mask the product. Regenerate only the background. Most tooling supports this; it is the single highest-value habit.
3. Never let text be regenerated. If packaging copy is visible, either mask it or composite the original label back over the result.
4. Fix the colour afterwards. Sample the hex value from the original product photo and correct the output. Generated lighting drifts colour, and colour drives returns.
5. Review at the size customers see. A thumbnail hides artefacts that a product page does not.
6. Keep the original. When a customer asks whether the picture is real, you want to be able to show what the item looks like unstyled.

Where this fits in a small team

The realistic workflow for a shop with no studio: one honest photograph per product on a plain background, then generated variants for the channels that need atmosphere — a social post, an ad, a newsletter hero. The plain photograph stays on the product page, where accuracy matters most and where returns are decided.

That split also keeps the disclosure question simple. The product page shows a photograph. The campaign creative is generated, labelled and never claims to be documentation.