On this page
What Article 50 asks for
The transparency rules are in Article 50 of the AI Act. For images, two paragraphs matter, and they address different parties.
- Providers (Article 50(2)). If you develop a system that generates synthetic audio, images, video or text and offer it under your own name, its output must be marked in a machine-readable format and be detectable as artificially generated or manipulated. The technical solution should be effective, interoperable, robust and reliable, as far as technically feasible.
- Deployers (Article 50(4)). If you use such a system professionally to publish a deepfake (content that resembles existing persons, objects, places or events and would falsely appear authentic), you must disclose that it was generated or manipulated. For evidently artistic or creative work, an appropriate disclosure that doesn't spoil the work is enough.
There are exceptions. The marking duty doesn't apply where a system only performs an assistive function for standard editing, or doesn't substantially alter the input. Virtual staging and product restyling clearly do alter the image, so we treat every render as in scope.
A virtually staged photo of a real home arguably fits the deepfake definition once it's published. That's one more reason we give the agents and sellers who publish our renders a visible disclosure, not just metadata.
When it applies
Article 50 applies from 2 August 2026. The Digital Omnibus on AI, Regulation (EU) 2026/1744, left the disclosure duties in place but added one transition: providers of generative systems that were already on the market before 2 August 2026 have until 2 December 2026 to comply with the machine-readable marking in Article 50(2).
Content generated before 2 August 2026 doesn't need to be labelled retroactively, according to the Commission, which still encourages it where possible.
For practical guidance there is the Commission's Code of Practice on Transparency of AI-Generated Content, published in June 2026. It's voluntary, and it asks for machine-readable solutions that are effective, interoperable, robust and reliable, as far as technically feasible.
Machine-readable marking with IPTC metadata
The most widely supported machine-readable signal for images today is the IPTC DigitalSourceType property, which several large platforms read to recognise AI imagery. The IPTC vocabulary separates images created from scratch from photos edited with generative AI:
| Term | IPTC name | Use it for |
|---|---|---|
trainedAlgorithmicMedia | Created using Generative AI | Images generated from a prompt, with no source photo |
compositeWithTrainedAlgorithmicMedia | Edited using Generative AI | Real photos changed with generative AI, such as inpainting, staging or restyling |
algorithmicallyEnhanced | Algorithmically-altered media | Corrections like sharpening or noise reduction that don't change the content |
Our renders start from a real photo (an empty room, a product on a kitchen counter), so we tag them as compositeWithTrainedAlgorithmicMedia. Next to it we write three human-readable tags, so anyone who inspects the file can see where it came from:
DIGITAL_SOURCE_TYPE =
"http://cv.iptc.org/newscodes/digitalsourcetype/compositeWithTrainedAlgorithmicMedia"
AI_DISCLOSURE_ARGS = [
"-Credit=Made with twelvelumens.com",
"-Creator=Created using Twelve Lumens generative AI",
"-Software=Twelve Lumens (twelvelumens.com)",
"-XMP-iptcExt:DigitalSourceType=#{DIGITAL_SOURCE_TYPE}"
].freeze
# Uploads: remove everything except the colour profile.
def strip!(path)
run!("-m", "-overwrite_original", "-all=", "--icc_profile:all", path)
end
# AI renders: remove everything except the colour profile, then write the disclosure tags.
def stamp_ai!(path)
run!("-m", "-overwrite_original", "-all=", "--icc_profile:all", *AI_DISCLOSURE_ARGS, path)
end Stripping first matters: -all= removes whatever metadata the model or an earlier step left behind, --icc_profile:all keeps the colour profile, and only then do we write our tags. Renders are stamped before they're stored, so what leaves the system carries the label. To check a file, read the tags back:
exiftool -XMP-iptcExt:DigitalSourceType -Credit -Creator -Software render.jpg Visible labels that survive a screenshot
Metadata is invisible to the people looking at the picture, and it's easily lost: take a screenshot, re-encode the file or upload it to a site that strips metadata, and the label is gone. For anything published to buyers, we add a disclosure people can see.
Both products burn a full-width bar into the bottom of the image with libvips, carrying a short disclaimer such as ‘Digitally styled for inspiration, may slightly differ from reality’. A full-width bar is harder to crop away than a corner badge, and it scales with the image, so 1K and 2K renders read the same.
- Mandatory on free plans. Free renders always carry the bar. Paid plans choose when to apply it.
- Your own wording. Customers on Pro and Business create disclaimer templates with their own text and logo.
- Five languages. Templates are translated by AI into English, Dutch, German, French and Spanish, so the disclosure matches the listing.
- Labelled twice. The barred output gets the same IPTC tags, so the visible and the machine-readable label travel together.
def call(input_path, output_path)
image = load_image(input_path)
bar = build_bar(image.width, bar_height(image)) # full width, scales with the image
marked = image.composite2(bar, :over, x: 0, y: image.height - bar.height)
.flatten(background: [255, 255, 255])
Images::Encoder.write(marked, output_path, format: @format)
Images::Metadata.stamp_ai!(output_path) # visible and machine-readable
output_path
end
def bar_height(image)
(image.height * 0.052).round.clamp(34, 120)
end One detail worth copying: the brand credit on the right gets its own share of the width, so a long credit can never shrink the disclosure on the left, which is the part that actually matters.
The file format trap
How well metadata survives differs per file format and per tool. exiftool silently drops our IPTC Credit tag when writing WebP. The file looks fine and the export succeeds, but part of the label is gone.
So our exports only offer JPG and PNG. Renders are stored as lossless PNG masters and delivered as JPG by default, because listing portals expect it. The export path stamps the file after encoding. WebP and AVIF stay out until we can show that every tag survives.
- Test every format you export with
exiftooland check every tag, not just the AI one. - Stamp after the last encode or resize. Any later step that rewrites the file can strip the tags.
- Watch your image pipeline: CDNs and resizing services can strip metadata from the images they deliver, so keep a visible label on anything you publish.
Strip metadata on the way in
Labelling outputs is half the job. The other half is not leaking what comes in. Phone photos carry GPS coordinates, device serial numbers and editing history. Both products strip third-party metadata from every upload before it reaches storage, keep only the colour profile, and convert HEIC to JPEG on the way.
Our vision pass also notes whether an upload already shows a visible AI label, such as ‘AI-generated’ or ‘virtually staged’, so you know what you're working with before anything is rendered.
Where C2PA fits
C2PA Content Credentials go a step further than IPTC tags: a cryptographically signed manifest that records how an image was made and edited. Like any metadata it can be lost when a file is screenshotted or re-encoded, which is why it's usually paired with visible labels or watermarks.
We don't sign C2PA manifests today: our renders carry IPTC tags, plus a visible bar on free plans and wherever customers turn it on. If your images travel through many systems, check the Code of Practice for the techniques it expects, and consider signed credentials or an invisible watermark as extra layers.
A checklist for your own product
- 1
Work out your role
Are you the provider of the generating system, a deployer publishing its output, or both? The duties differ.
- 2
Tag every output
Write the right IPTC
DigitalSourceTypeterm in the last step before storage or delivery, after stripping old metadata. - 3
Add a visible label where people see the image
Especially for realistic images of real places, people or products.
- 4
Test every export format
Read the tags back with exiftool for JPG, PNG and anything else you offer.
- 5
Strip what comes in
Remove GPS, device and editing metadata from uploads before you store them.
- 6
Write it down
Document what you mark, how and why, so you can show it when a customer, a platform or a regulator asks.
Would you rather have this built into your product than bolted on later? It's part of how we build AI photo editing software and generative AI apps.
Frequently asked questions
Does every AI-edited image need a label under the EU AI Act?
Not every one. The machine-readable marking duty in Article 50(2) applies to providers of generative systems, with exceptions for assistive standard editing and edits that don't substantially alter the input. Deployers who publish deepfakes have their own disclosure duty. Where your images fall depends on your system and how it's used, so check it with a lawyer.
Is IPTC metadata enough to comply?
It's a widely supported machine-readable mark, but the AI Act asks for solutions that are effective, interoperable, robust and reliable as far as technically feasible, and metadata is easily lost. Many teams pair it with a visible label, signed C2PA credentials or an invisible watermark.
When do the rules apply?
Article 50 applies from 2 August 2026. Providers of generative systems that were already on the market before that date have until 2 December 2026 to meet the machine-readable marking requirement.
Why not export WebP or AVIF?
Because the labels have to survive the export. In our pipeline exiftool silently drops the IPTC Credit tag when writing WebP, so we only offer formats where every tag demonstrably survives.
Can you add AI labelling to our product?
Yes. Both of our image products label every render this way, and we build the same pipeline into client products, for example in e-commerce platforms with AI product photos.