Labelling AI images in a cosmetics and beauty store: before-and-after, skin, textures, models

In a cosmetics or beauty store, the image often shows exactly what the product promises: smoother skin, a particular shade, a texture, a result. That puts AI images closer to the line drawn by Article 50(4) of the AI Act (Regulation (EU) 2024/1689) than in many other sectors. This guide sorts the typical uses of AI, from skin retouching and AI models to swatches and shade variants, following the European Commission’s guidelines of 20 July 2026 (C(2026) 5054 final) and the Code of Practice of 10 June 2026. It also shows where cosmetics law requires a separate check that no AI label replaces. The general table for all image types is in the guide Which images in your store need an AI label?

Which benchmark applies to beauty images?

If you use AI tools for store images, you are a deployer under Article 3(4) of the AI Act. For images, deployers are bound by Article 50(4): a deepfake must be visibly disclosed as artificially generated or manipulated, under Article 50(5) in a clear and distinguishable manner, at the latest when someone first sees the image. The machine-readable marking under Article 50(2) is the responsibility of the tool’s provider, not yours. The Digital Omnibus (Regulation (EU) 2026/1744) amended only Article 50(7) (Article 1 point 20); paragraphs 1 to 6 and the deepfake definition are unchanged, and Article 50 has applied since 2 August 2026. For the wider picture, see the guide on EU AI Act labelling requirements.

Under Article 3(60) of the AI Act, a deepfake is AI-generated or manipulated image content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful. Three clarifications in the guidelines decide most beauty cases:

• AI models are persons. Under paragraph 113, “persons” includes realistic AI-generated human avatars and personas, and “objects” expressly includes consumer goods. Authenticity there also covers “the actual appearance or use of objects”.
• The test is objective. No intent to deceive is required (paragraph 114). Photorealism makes a deepfake more likely but is not decisive on its own.
• The whole audience counts. If an image can foreseeably reach children, older people or people with lower AI literacy, it may be enough under paragraph 115 that it deceives that part of the audience.

When is an AI edit too small for a label?

Three exceptions are easily confused in the beauty sector:

1. Minor editing (paragraph 116). AI editing of insignificant details usually does not turn a real photo into a deepfake: lighting adjustments, colour correction, noise reduction, a background replaced for aesthetic purposes, arranging existing products. The guidelines also list “cosmetic adjustments and enhancements” there; this means superficial adjustments, not the cosmetics industry. Paragraph 116 makes this expressly subject to context and to the impact on perception in the specific case.
2. Assistive function for standard editing (Article 50(2), third sentence). This exception concerns the providers’ marking obligation, not your label. Its example lists after paragraph 92 still show where the Commission draws the line: removing dust spots and red-eye and minor colour and brightness corrections count as standard; substantial facial modification and altering a person’s body shape or skin colour do not.
3. Evidently artistic works (Article 50(4), third sentence, paragraphs 119 to 124). The exemption does not remove the label, it only allows a more discreet form (paragraph 123). It is to be interpreted strictly; recognisably commercial content falls outside it, and advertising falls within it only in specific situations (paragraph 122). As counter-examples the guidelines name simulated consumers demonstrating a product and a realistic synthetic influencer testing a product (examples after paragraph 124).

This leads to a rule of thumb for beauty (interpretation): if the AI edit concerns exactly what the product promises, meaning skin, complexion, lips, colour, texture or result, it is no longer insignificant. If it only changes the surroundings of an unchanged product, paragraph 116 usually applies.

Which beauty images need a label?

The overview follows the example lists in the guidelines; anything beyond them is flagged as interpretation or uncertain.

Skin and retouching: when does editing become a deepfake?

Retouching is routine in beauty, and paragraph 116 expressly excludes small edits. What matters is what the edit concerns:

• An image with no promise about skin. If the AI removes a dust spot, a stray strand of hair or a single blemish in a fragrance image, the image usually remains a real photo with standard editing. No label (interpretation; not officially decided).
• An image with a promise about skin. For a serum, foundation, primer or anti-wrinkle cream, the skin in the image is the statement about the product. If the AI smooths wrinkles, closes pores or evens out the complexion, it changes exactly what the viewer takes to be the result. On the reading taken here, such an edit is no longer insignificant within the meaning of paragraph 116: label AI MODIFIED (uncertain, as of 30 September 2026).
• Face and body. The guidelines name substantial facial modification and altering a person’s body shape or skin colour as edits beyond standard editing in the examples after paragraph 92; on the reading taken here, a slimmer face or fuller lips fall under this. That list belongs to the providers’ obligation under Article 50(2); for your label it is a strong indication, not a decision. Label AI MODIFIED.

Article 50 covers only AI editing. Manual retouching needs no AI label but remains subject to cosmetics law (see below).

Models and faces

• AI model from a prompt showing lipstick, mascara or a cream: a deepfake under paragraph 113, label AI GENERATED. The artistic exemption does not help; see the counter-examples after paragraph 124.
• Real product photo placed in a scene with an AI model: label AI MODIFIED, because a source photo was used; when in doubt, AI GENERATED.
• Real model, AI applies a make-up look or a lipstick colour: label AI MODIFIED. The image shows a result that was not photographed as such; if the colour is not exact, misleading advertising comes on top.
• If an AI face resembles a real, recognisable person, the right to one’s own image also applies (in Germany Sections 22 and 23 KUG; Austria and Switzerland have their own rules). Check generated faces for resemblance to well-known people.

Before-and-after images

Before-and-after images only work because both images are real. That is exactly why AI versions regularly meet the fourth deepfake criterion: they would falsely appear authentic.

An AI-enhanced after image with a label is still an image showing an effect the product did not achieve in that form. The fix is not the label but a real photo. Whether a labelled, fully AI-generated image may serve under cosmetics law as a mere illustration of a substantiated effect is unsettled (uncertain, as of 30 September 2026).

Textures, swatches and shade variants

For make-up and nail polish, the shade is a central product characteristic; for skincare, the texture. Both are statements about the product itself:

• AI-generated texture or swatch images (a smear of cream, pigment on skin, polish on a nail) show a consumer good. If the depiction deviates, it is a deepfake under the positive list after paragraph 116; if it is accurate, it remains a borderline case like the fully generated product photo. When in doubt, AI GENERATED.
• Shade variants by AI: from a photo of shade 01, the AI creates shades 02 to 12. It thereby alters the product itself; under the rule of thumb in the general guide, that needs at least AI MODIFIED. If a shade does not match the real colour, the image shows a product that does not exist in this form: that is also misleading (Section 5 UWG in Germany, Section 2 UWG in Austria) and, under Article 20(1) of Regulation (EC) No 1223/2009, a picture implying a characteristic.
• The safest route: photograph every shade and use AI only for cutouts and for colour correction against a colour reference.

Packshots and campaign images

• Real product photo on an AI background (marble, water, silk): no label. This is the case from the negative list after paragraph 116, as long as the background does not mislead about the product; paragraph 116 also names AI arrangements of existing products.
• Ingredients in the background: if the AI shows fruit, flowers, plants or droplets, the image makes a statement about ingredients or effect. Under Article 1 of Regulation (EU) No 655/2013, pictures are claims too, and under point 2(1) of its Annex, an ingredient that is claimed must be deliberately present in the product. If the background misleads about characteristics, the negative list no longer applies either. For food, Article 7 of the FIC Regulation applies, see AI images in food and drink stores.
• Fully AI-generated packshot: if it shows the jar more appealing or different from what is delivered, label AI GENERATED; if it shows it accurately, borderline, label when in doubt.
• Clearly surreal campaign images: a lipstick as tall as a skyscraper or a cream jar as an island in the sea resemble no plausible reality and do not mislead (paragraph 113, criterion ii). No label. If a photorealistic model appears in the image, the section on models applies to that person.

The cosmetics factor: what an AI label does not settle

For cosmetic products, EU cosmetics law applies alongside the AI Act. It does not ask whether AI was involved, but whether the image’s statement is true and substantiated. The two checks run separately; an AI label answers only the first.

• Article 20(1) of Regulation (EC) No 1223/2009: in the labelling, making available on the market and advertising of cosmetic products, “text, names, trade marks, pictures and figurative or other signs shall not be used to imply that these products have characteristics or functions which they do not have”. Pictures are named expressly.
• Regulation (EU) No 655/2013 sets the common criteria for claims and, under Article 1, also covers pictures that convey product characteristics or functions “explicitly or implicitly”. Most relevant for images: claims must be supported by adequate and verifiable evidence (Annex, point 3(1)), and presentations of a product’s performance must not go beyond the available supporting evidence (Annex, point 4(1)). Statements of clear exaggeration that no one takes literally need no substantiation (Annex, point 3(5)).
• Who ensures this: under Article 2 of Regulation (EU) No 655/2013, the responsible person within the meaning of Article 4 of Regulation (EC) No 1223/2009 ensures that claims meet these criteria and are consistent with the documentation of the claimed effect in the product information file (Article 11(2)(d)). If you place a cosmetic product on the market under your own name or trade mark, you are the responsible person yourself under Article 4(6).

For AI images this means: an AI-enhanced result is no substantiation of an effect. The label discloses that AI was involved; it does not make an unsubstantiated presentation of an effect permissible. Whether a specific effect claim is substantiated is a separate topic, answered by the product information file, not by the image label. This section describes the legal situation and is not legal advice.

On scope: both regulations apply directly in all EU Member States, including Germany and Austria. Switzerland has its own rules, and products that are not cosmetics but, for instance, medicinal products or medical devices are subject to their own advertising rules (in Germany the Heilmittelwerbegesetz). This guide covers neither.

AI GENERATED or AI MODIFIED?

Annex 1 of the Code and the guidelines (paragraphs 58, 59) decide between the two worded icons with one question: was there a source photo?

If in doubt between the two, choose AI GENERATED, the stronger disclosure. Under Section 2, sub-measure 1.2.2(a) of the Code, the label belongs in a place without intervening overlay elements, for example the top right corner; if a “new” or discount badge already sits there, it must not cover the label. Add a text alternative (alt text or an aria-label on the label), because the second sentence of Article 50(5) refers to accessibility requirements. Ads, social media and marketplaces need a derivative with a burned-in label; more on this in the guide Which images in your store need an AI label? What the platforms require is covered in the guide on AI images on marketplaces and in ads.

What do provenance markers reveal about beauty images?

Many generators write a provenance marker into the file on saving (C2PA, IPTC DigitalSourceType). A found marker is evidence of AI origin; a missing marker is no statement. Three findings from the Easy AI Act Image Labels test bench matter for beauty images:

• The marker can separate generated from edited, but not required from optional. On 1 September 2026, a Canva image made from one’s own photo with an AI background carried the value compositeWithTrainedAlgorithmicMedia, a Canva image from a prompt only trainedAlgorithmicMedia. Yet the typical packshot on an AI background usually needs no label under paragraph 116: the marker says that AI was involved, not whether a label is required.
• Many edits leave no marker. The Shopify image editor’s “Generate background” wrote not a single metadata block on the same day, and 13 of 47 models in one generator catalogue set no marker. Tools that only retouch skin have not yet been measured by the test bench. Skin smoothed by AI without a marker may therefore still need a label.
• The store copy is not fit for checking. On the copies Shopify delivers to visitors, only 6 of 38 marked images still carried the marker on 30 September 2026; markers that sat only in the C2PA manifest survived no re-encoded copy (0 of 31). So check the original file, for example in the AI image checker. Which tools write what is shown in the guide Which AI image generators write provenance evidence?

Checklist for a beauty store

1. Sort images by type: packshots, textures and swatches, shade variants, model and campaign images, before-and-after.
2. Answer two questions per image: was there a real source photo? And did the AI change only the surroundings, or the product, skin, face or result?
3. Classify retouches: if the retouch shows the advertised effect, the image gets AI MODIFIED; small retouches unrelated to the product promise usually need no label.
4. Label AI models: from a prompt AI GENERATED, with a real product photo as the source AI MODIFIED; check generated faces for resemblance to real people.
5. Before-and-after with real photos only: same conditions and same editing on both sides; replace AI-enhanced after images with real photos, a label does not make them permissible.
6. Match colours: check swatches and shades against the real product, photograph every shade where possible; shade variants recoloured by AI get at least AI MODIFIED.
7. Check backgrounds: show ingredients, fruit or plants only if the ingredient is deliberately present.
8. Match effect claims: whatever effect an image shows must be consistent with the documentation in the product information file; for your own brand, that is your responsibility.
9. Place the label visibly: top right without intervening elements, with a text alternative for screen readers; for ads, social media and marketplaces, a derivative with a burned-in label.
10. Record decisions: note the tool, source photo and decision per image, including when no label is set.

An app like Easy AI Act Image Labels reads provenance markers on the original and proposes labels. Whether an image without a marker needs a label, and whether the presentation of an effect holds up under cosmetics law, remains your decision.