Labelling AI images in a fashion store: models, virtual try-on, backgrounds

In a fashion store, the AI label mostly comes down to one question: does the image show a person who does not exist in this form, or a garment as it was never photographed? Photorealistic AI models, including virtual try-on with your real product photo, need a visible disclosure under Article 50(4) of the EU AI Act (Regulation (EU) 2024/1689). Real photos that were only retouched, cut out or placed in front of a designed background usually need no label, as long as the model and the garment stay unchanged. In between lie borderline cases such as AI colour variants, generative ghost mannequins and altered body shapes. The European Commission’s guidelines of 20 July 2026 (C(2026) 5054 final) do not mention fashion, clothing or virtual try-on expressly; the classification below follows their general criteria and examples and flags where it is an interpretation of its own. For the full picture, see the Article 50 overview.

When is a fashion image a deepfake?

The obligation falls on you as a deployer: if you use a finished AI tool in the course of business, you are a deployer under Article 3(4) of the AI Act and, under Article 50(4), must disclose that a deepfake has been artificially generated or manipulated. The machine-readable marking in the file under Article 50(2) is owed by the provider of the tool. You remain the deployer even if an agency or a photo studio creates the images, as long as you decide whether and how AI is used (guidelines para. 14).

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. The guidelines break this down into four cumulative criteria (para. 113). For fashion images, three clarifications matter:

• Persons are realistic human beings, expressly including realistic AI-generated human avatars and personas and digital replicas of real persons. A photorealistic AI model does not have to resemble anyone; it is enough that it could plausibly exist.
• Objects are realistic material items, expressly including consumer goods, and events are realistic scenes or situations that can involve persons. Under para. 113, authenticity also covers the actual appearance and use of objects.
• The test is objective, based on context and the expected audience, and does not require any intent to deceive (para. 114). A high degree of photorealism makes a deepfake more likely but is not decisive on its own.

The artistic exemption hardly helps in a store. If a deepfake forms part of an evidently artistic, creative, satirical or fictional work, the obligation shrinks to a disclosure that does not hamper the enjoyment of the work (Article 50(4), first subparagraph, third sentence); it does not disappear (para. 123). The categories are to be interpreted strictly, recognisably commercial content falls outside them, and where content is mixed, the informative character prevails (para. 122). As examples without the exemption, the guidelines name a realistic synthetic influencer testing out a sponsored real product and simulated consumers in the style of a teleshopping channel. A visibly surreal motif, such as a model floating without any aid, may fall outside the deepfake definition altogether (para. 113) as long as it cannot mislead about the garment; it then rarely works as a product image.

Minor editing does not create a deepfake. Under para. 116, AI edits of insignificant details may be of minor relevance to whether content appears authentic. The paragraph names, among others, colour correction, lighting adjustments, noise reduction, cosmetic adjustments, editing background details and, in product advertising, replacing backgrounds for clearly aesthetic purposes, each subject to the specific case. Article 50(2) draws a second line for providers: tools that only perform an assistive function for standard editing or do not substantially alter the input need not mark their output. The examples in para. 92 are telling for fashion images. Deleting backgrounds that are visible in the original counts as standard editing; face replacement, altering a person’s body shape or skin colour, replacing objects or persons in a way that changes the image’s meaning, and composites that modify the representation of persons or objects do not. These examples apply directly to the provider’s obligation; for your disclosure under paragraph 4 they are an indication (interpretation).

Which AI models and images of people need a label?

With AI models the situation is therefore clearer than with product images: the guidelines name realistic AI avatars expressly, and the examples without the artistic exemption come from advertising.

What applies to backgrounds, retouching, cut-outs, ghost mannequins and colour variants?

The rule of thumb from the other guides applies here too: if the AI changes how the garment or the person wearing it looks, the image needs at least AI MODIFIED. If it only changes the surroundings of a piece shown unchanged, para. 116 usually applies. For the borderline cases, as of 31 August 2026 there was no court decision in Germany, Austria or Switzerland. The Wettbewerbszentrale (the German unfair-competition association; guidance version 2 of 29 July 2026) and the Austrian Schutzverband gegen unlauteren Wettbewerb (position of 4 February 2026) take the broad reading: label photorealistic AI images when in doubt. Edits made without an AI tool, such as a ghost mannequin assembled by hand, are not touched by Article 50 at all.

AI GENERATED or AI MODIFIED?

The European Commission provides three icons: the basic “AI” icon, “AI GENERATED” and “AI MODIFIED”. On its icon page (as of 24 September 2026), it assigns them as follows: AI GENERATED when the content is fully generated by AI with no human-created elements apart from prompting; AI MODIFIED when pre-existing, human-made content was partially modified with AI, turning it into a deepfake. This matches the guidelines, which distinguish generation from a prompt and manipulation of existing content (paras. 58, 59). So the question is: is there a real photo in the image?

In virtual try-on, the model, the pose and often the scene are invented, while the garment comes from your photo; under the Commission’s assignment, that is AI MODIFIED. If you do not know whether a tool used your photo or merely recreated the piece, choose AI GENERATED, the stronger disclosure. The basic “AI” icon on its own is ambiguous: according to Annex 1 of the Code of Practice of 10 June 2026, the variants with a text label performed significantly better in user tests. If there is no explanatory text right next to it, the word belongs in the label.

What is special about fashion?

• The fit in the image is computed. An AI model has no body the piece was ever worn on; in virtual try-on, the tool calculates drape, length and width from a photo. The guidelines’ list of deepfakes expressly includes an AI product image that can mislead as to the product’s appearance, characteristics or use, for example by making the product appear more appealing or of better quality than in real life. In fashion, how a garment looks on the body is part of the product’s appearance (interpretation).
• A label does not cure misleading advertising. Article 50(6) of the AI Act leaves other transparency obligations untouched, and unfair competition law assesses the image independently of the label. In Germany, Section 5(2) no. 1 UWG expressly lists type (Art), design (Ausführung) and condition (Beschaffenheit) among the essential characteristics of goods, and a product image is a statement about them. If an AI image shows the garment differently from what you deliver, for example in colour, material or cut, the image itself is the problem, labelled or not; in Austria, Section 2 of the Austrian UWG applies, in Switzerland Article 3(1)(b) of the Swiss UWG. The guidelines keep the two tests apart: the deepfake criterion of false appearance of authenticity is to be understood independently of the concept of deception in the Unfair Commercial Practices Directive (footnote 32).
• “Image may differ” is not a label. The phrase does not say that an image comes from AI, and it points to the real problem: if the product differs, the difference is the problem, not the AI. The same goes for “image for illustration purposes only”.
• Real people, real rights. A digital twin, or an AI model that recognisably resembles a real person, also touches personality rights; what counts is recognisability, not intent. In Germany, Section 22 KUG allows portraits to be distributed only with the consent of the person shown, subject to exceptions; in Switzerland, the right to one’s own image derives from Article 28 of the Civil Code (ZGB). Do not let AI tools create faces that resemble celebrities, and check the results.
• Young audiences. If your store also addresses teenagers or children, it may be enough that an image appears authentic to that part of the audience (para. 115).

Where does the label have to appear in a fashion store?

The obligation applies to each individual image and to every viewer, at the latest at first exposure (Article 50(5), para. 143). In a fashion store, customers often see an AI model first in the collection, not on the product page. Therefore:

• Every image location counts: collection grid, product gallery, variant images per colour, lookbook, blog and newsletter. If the collection only shows the model photo on mouse-over, the label has to appear with that image.
• Visible without a click: a badge on the image, preferably top right in a place with no intervening overlay elements under the Code of Practice (sub-measure 1.2.2(a)), or a caption clearly assigned to the image. The Code treats an overlay in the store that appears on the image as equivalent to a burned-in icon (sub-measure 1.2.1(c)). Make sure that sale, new-in or other badges do not cover the label.
• Not enough: a sentence inside a collapsed accordion, the footer or the terms and conditions (para. 142), metadata in the file (para. 117), a tooltip or alt text alone. Add a text alternative anyway (alt text or an aria-label on the label), because the second sentence of Article 50(5) refers to accessibility requirements.
• Social media, ads, marketplaces: an overlay from the theme does not travel with the image when it leaves the store. For social posts, ads and newsletters you need a derivative with a burned-in label; the original file stays without one. The platform rules are in the guide Which images in your store need an AI label? What the platforms require in detail is covered in the guide on AI images on marketplaces and in ads.

What do provenance markers show for fashion images, and what not?

Many AI tools write a provenance marker into the file when saving, either a C2PA manifest (Content Credentials) or the IPTC value DigitalSourceType. A found marker is evidence of AI origin; a missing marker is no statement. The measurements on the Easy AI Act Image Labels test bench show four things for fashion images:

1. Not every tool marks. On 1 September 2026, 34 of 47 text-to-image models in one generator catalogue (Magnific/Freepik) wrote a marker and 13 did not. An image from the Shopify image editor (“generate background”) carried no metadata at all in the stored original. The test bench has not measured specialised tools for AI models or virtual try-on; this guide makes no statement about them.
2. A marker can separate generated from edited, if the tool does so. For an image from a prompt alone, Canva wrote the value trainedAlgorithmicMedia (fully AI-generated); for its own photo with an AI background, compositeWithTrainedAlgorithmicMedia (in IPTC terms, augmentation, correction or enhancement using a generative AI model). What the image shows, whether an AI model or just a new background, and whether it needs a label, the marker does not say: it reflects what the creating tool declared.
3. The export path decides. The same generation carried a 3,174-byte manifest as a native PNG and no marker at all as a JPEG from the web interface; a screenshot of a marked file kept only 144 bytes of EXIF and 54 bytes of IPTC with no provenance information. So keep the original in its native output format before you edit or export images further.
4. The copy in the store usually loses it. The original stored in Shopify kept all 34 markers on 1 September 2026. On the copies Shopify delivers to visitors, only 6 of 38 marked images still carried the marker on 30 September 2026, and 5 in the product page srcset; markers held only in the C2PA manifest survived no re-encoded copy (0 of 31). Whether a fashion image contains AI therefore usually cannot be read from the delivered copy.

An app such as Easy AI Act Image Labels reads the markers on the original and proposes labels; sorting all images without a marker by the tables above remains your decision. More on detection methods in the guide Evidence or hint, and on Shopify’s own image editor in Shopify Magic and the EU AI Act.

Practical checklist for fashion stores

1. Take stock. For each group of images, list which tool was involved: AI model, virtual try-on, background, retouching, colour variant. Ask your agency, photo studio and team as well.
2. Classify each image with the two tables: label, no label or borderline case. Label photorealistic borderline cases.
3. Choose the label: no real photo in the image, AI GENERATED; real photo turned into a deepfake by AI, AI MODIFIED; unclear origin, AI GENERATED.
4. Check against the goods: do colour, material, length, width and drape match the piece you deliver? If not, change the image; a label does not change that.
5. Check the people: no celebrity-like faces; for digital twins and AI changes to real models, record the consent in writing.
6. Cover every image location: collection, mouse-over image, product gallery, variant images, lookbook, blog, newsletter; label visible without a click, not covered by other badges, with a text alternative for screen readers.
7. Serve the channels: for social media, ads and marketplaces a derivative with a burned-in label, the original without.
8. Keep the originals: download AI images from the tool in their native format, no screenshots, no conversion.
9. Document and agree: which image comes from which tool, which label, and why. If you have not signed the Code of Practice, you must be able to show how you meet the obligation (para. 148). Contracts with agencies and photo studios should state whether AI may be used, which images contain AI and who labels them.

The basics on roles, deadlines and the form of the label are in the guide Do I have to label AI images in my online store?