Label AI images

What imgmarker does

Upload an image and you get two markings back. One you can see: a badge in the corner that says the picture came from AI. One machines read: an IPTC field inside the file. You pick which of the two labels applies and which corner the badge sits in, then download the result. No account, no installation, up to ten images at a time.

The badge lives in the pixels, so it survives screenshots, re-uploads and platforms that throw away file information. The metadata sits in the file itself, where image software, search engines and content platforms find it without being told to look.

Your images travel to a server in Germany and no further. We delete them 60 minutes after processing, together with the database record - no archive and no backup. We do not analyse what your pictures show and never pass them to anyone. Within that hour we do review uploads to stop illegal material, which is the only reason anyone here looks at an image.

Why labelling matters

Article 50 of the EU AI Act sets out transparency duties for artificially generated content, applicable from 2 August 2026. It contains two obligations that people routinely mix up.

If you provide a generative AI system, you must mark its output so machines can detect it, through metadata, watermarks or provenance data. If you publish a deepfake, meaning content showing real people, places or events in a way viewers could mistake for reality, you must disclose that openly to those viewers.

Each marking on its own leaves a gap. A badge painted onto a picture tells a human reader everything and a machine nothing. Metadata tells a machine everything and a human nothing. imgmarker writes both in the same pass.

The law aside, labelling protects you. Newsrooms, agencies, stock libraries and marketplaces all ask where a picture came from now. An unlabelled AI image that later turns out to be synthetic costs you more credibility than one that said so from the start.

How to use it

1. 1

   Choose the label and position

   Above the drop zone you set which label applies and which corner the badge goes in. Those settings carry over to every image you add, so ten pictures labelled the same way cost you no extra clicks. Any single image can still be changed afterwards in the list.

2. 2

   Add your images

   Drag files onto the drop zone or pick them from your computer. JPG, PNG and WEBP work, up to 25 MB per file and 10 files per batch. Each row shows a thumbnail with the badge position marked, so you can see where it will land before you start.

3. 3

   Add details, if you want

   Under “Advanced details” you can supply a creator, a description and a credit line. These go into the file as standard IPTC fields, the same ones photo agencies and editorial systems already read. They apply to every image in the batch.

4. 4

   Start and download

   Each image runs on its own, so one broken file never holds up the rest. Take finished pictures one at a time or grab the whole batch as a ZIP. Your filename stays as it was, with “-aimarked” added before the extension.

What exactly goes into the file

The value that matters is the IPTC field DigitalSourceType, stored in the image’s XMP packet. Google, Adobe and the major content platforms read this vocabulary, which makes it the closest thing to a shared standard for saying where a picture came from.

Alongside it goes a plain sentence in the description field, plus whatever creator, credit and description you supplied. JPEG files get the same sentence in the older IPTC-IIM and EXIF fields as well, so software too old for XMP still shows something useful.

Check the result yourself if you like. Any metadata viewer displays the field, and on the command line exiftool prints it in one line.

Why both markings

Metadata is exact and easy to lose. Several large platforms strip it during upload, sometimes to save space, sometimes to remove location data. Once that happens the machine-readable marking is gone and the badge is all that is left.

The badge holds up where metadata fails. It comes through screenshots and re-encoding intact. What it cannot do is announce itself to software: no automated system tells it apart from any other text sitting in the picture.

One marking fails where the other survives. Use both and an image usually arrives with at least one of them still attached, whatever route it took to get there.

A note on Content Credentials

Some images arrive carrying C2PA Content Credentials, a signed record of where the picture came from and what was done to it. Adding a badge breaks that record. The signature covers the original pixels, and drawing on the image changes them.

imgmarker spots this before processing and asks you to confirm. Once you do, we strip the credentials instead of leaving them in place, because verification tools read a broken signature as evidence of tampering. Users end up worse off than if the picture had carried no credentials at all.

If keeping those credentials matters to you, label the file somewhere else. An editor that understands C2PA can add the badge as a signed edit and keep the chain intact.

What this tool will not do

imgmarker cannot tell you whether an image came from AI. You decide which label fits and the tool applies it. Nobody has a reliable detector for AI-generated pictures, so treat any service that claims one with suspicion.

Nor is this legal advice. Whether your label satisfies the rules depends on your role, your content and where you publish. If the answer carries real consequences for you, ask a lawyer.

And no marking survives someone determined to remove it. A badge can be cropped out, metadata can be deleted. Labelling is how you stay straight with your audience, not how you stop bad actors.

Frequently asked questions