Skip to main content
Demarker

Provenance cleanup for your own AI images

See what your AI images carry. Then clean them.

A free scan shows visible logos, AI metadata and the hidden watermarks it can infer. Then clean your own images, free up to 3 a day.

  • Originals deleted 24 h after your last run finishes, or after upload if you never process (up to 72 h while awaiting payment)
  • Never used for training
  • Credits back on failure

The scan reads the marks and metadata these generators write, when the file still has them

  • ChatGPT
  • Gemini
  • Nano Banana
  • Copilot
  • Grok
  • Midjourney
  • Stable Diffusion

Not affiliated with or endorsed by any generator. Use on images you own or have the right to edit.

What is hidden inside

The marks you cannot see.

An AI image can carry its origin in three places. Each one needs a different kind of cleanup.

  1. Layer 1

    Visible marks

    Painted into a corner

    Logos and sparkles that some generators stamp onto the image. You can see them. We find registered marks by name and paint them out.

  2. Layer 2

    AI metadata

    Stored with the file

    Content Credentials (C2PA) and EXIF, XMP and PNG text tags that say how the image was made. They are signed and they travel with the file. Stripping them is the easy part.

  3. Layer 3

    Hidden in the pixels

    Woven into the image data

    SynthID and similar invisible watermarks. There is no field to delete, and we cannot decode them. We infer them from the metadata.

Metadata strips are easy, and many free tools do them. Pixel watermarks are not. We handle visible marks and metadata. For hidden watermarks we make a best-effort attempt, and we cannot confirm the result from the file. We tell you which is which before you pay for anything.

The report

What a scan reads in your file.

The scan is free. It lists each signal it finds and what each clean would do with it.

scan-reportExample
SignalWhere it livesQuick cleanGPU pass
Visible logo or sparkleCorner of the imagePainted outRegistered marks onlyPainted out
Content Credentials (C2PA)File metadataRemovedRemoved
AI tags (EXIF, XMP, PNG text)File metadataRemovedRemoved
Author and copyright fieldsFile metadataKeptNot confirmed
Hidden watermark (SynthID)Image pixelsNot touchedDisruptedBest effort

Hidden watermarks cannot be decoded here. Removal is best effort and cannot be confirmed from the file. Author and copyright fields are kept by Quick clean. For the GPU pass that is not confirmed yet.

Example only. This is an illustration, not the result for your file. Your own scan lists what your file actually carries.

Example

One visible mark, before and after.

An illustration made for this page, with a plain corner badge standing in for a visible mark. Results depend on your file: some marks leave faint traces, and Quick clean does not touch a hidden watermark.

Example, not your file

The same illustrated landscape with the badge painted out
An illustrated landscape with a small outlined square badge in the bottom right corner
BeforeAfter
Illustration, 1600 × 1000 pxDrag to compare
  • 24h

    Originals deleted after your last run finishes

  • 72h

    The longest an original waits while a job awaits payment

  • 3/ day

    Free Quick cleans on the web

  • 10credits

    For one GPU pass, output up to 12 MP

Scan an image

The scan is free. For images you own or have the right to edit.

Ownership

Your images. Your file.

You generated it. We remove the AI provenance tags and keep any author and copyright fields that are already in your file.

Read the Terms
  • Your author and copyright details stay in the file by default.
  • We refuse stock, marketplace and third-party marks, and adult content.
  • You confirm the image is yours before each clean.

How it works

Four steps. The first two are free.

  1. 1

    Drop your image

    PNG, JPG, WebP or AVIF, up to 25 MP. It is uploaded to our servers and, for the GPU pass, to a GPU host we rent. It is not processed in your browser.

  2. 2

    We scan three layers

    Free

    Visible marks, AI metadata, and hidden watermarks inferred from the metadata. You see the report before you choose anything.

  3. 3

    Choose a clean

    Quick clean: free, 3 a day

    Quick clean paints out registered visible marks and strips AI metadata, with output up to 12 MP. Larger images are downscaled to 12 MP unless you pay for original size. The GPU pass redraws the image to disrupt hidden watermarks, and uses credits.

  4. 4

    Check and download

    Slide between before and after, see what was removed and what stayed, then download. If a GPU pass did not work, re-run it free within 23 hours.

Under the hood

What happens to your file.

  • Retention

    24h

    After your last run

    Originals deleted in 24 hours

    Your original is deleted 24 hours after the last run on it finishes, or after upload if you never process it. It can stay up to 72 hours while a job awaits payment. We never use images for training, and you can delete an image and its results at any time with Delete now.

  • Your file

    Never

    Used for training

    Your author details stay

    Creator and copyright fields stay in the file by default. We strip AI provenance tags only. If a Content Credential carries author values, we copy them into the standard fields first.

  • If it fails

    Credits back

    If processing fails

    A free re-run if it did not work

    If a GPU pass did not work, tell us and re-run it once, free, within 23 hours. If processing fails, your credits come back.

Who it is for

Made for people who generate their own images.

  • Publishing

    You publish your own AI work

    Keep your name and copyright in the file and take the generator's tags out. Say how the work was made wherever the law or your client expects it.

  • Archive

    You keep an archive

    Make a clean copy of each image you generated. Your original file is never overwritten, so you can always go back to it.

  • API

    You build with it

    Upload, scan, clean, webhook. The scan is free over the API too.

    API overview

Pricing

Simple pricing, no subscription

One credit is one Quick clean with output up to 12 MP. The scan is free, and so are 3 Quick cleans a day on the web. Credits come back automatically if processing fails.

Free

No payment needed.

  • Free scans (fair-use limits apply)
  • 3 Quick cleans a day, output up to 12 MP. Larger images are downscaled to 12 MP unless you pay 2 credits for original size.
Scan an image

Credits

One-time packs, below.

  • The GPU pass
  • Large images, up to 25 MP
  • The API
  • Credits never expire
See packs
Quick clean
1 credit, or 2 for the original size above 12 MP
GPU pass
10 credits, or 15 for the original size above 12 MP

One-time packs, credits never expire

Standard

Recommended

110credits

$12.00

$1.09 per GPU image

Buy Standard

Pro

Best value

300credits

$30.00

$1.00 per GPU image

Buy Pro

Rates are per GPU image, with output up to 12 MP. The GPU pass is best effort. It redraws the image, faces can change slightly, and we cannot confirm removal from the file. See Limitations.

Limitations

What we cannot promise.

Read this before you pay for a GPU pass.

  • Removal of hidden watermarks is best effort

    Detectors change over time. A mark we disrupt today may still be read by a newer checker, and we cannot tell you which.

  • A clean image can still be recognised as AI

    Other methods, such as classifiers and a trained eye, do not depend on a watermark or on metadata.

  • The GPU pass redraws the image

    Faces can change slightly. Compare before and after, and keep your original.

  • We cannot confirm success from the file

    We cannot decode hidden watermarks, so we cannot check the result from the file. Keep your original.

  • Labelling rules still apply to you

    Some places require AI-generated or AI-edited content to be labelled. Removing a mark from a file does not remove that duty, and you are responsible for it.

Learn

What the marks are, in plain terms.

What is SynthID?

SynthID is a watermarking technology from Google DeepMind. It adds a mark that people cannot see to content made with some Google AI tools. For images, the mark is part of the pixel values themselves, so it is not a metadata field that can be read or edited like a caption.

It is designed to survive common edits such as resizing, compression and filters, and a matching detector can read it. Google controls that detector. This site does not decode SynthID, so a scan here can only tell you whether the file's metadata points to a generator that is known to watermark.

Other companies use their own invisible watermarks, and they work in a similar way. Our report names a hidden watermark only when the file's metadata points to a generator that is known to use one. That is an inference, not a detection. If you need to know whether an image carries SynthID, use Google's own checking tools.

What a GPU pass can and cannot do about it

SynthID vs C2PA Content Credentials

C2PA is an open standard for provenance. Content Credentials is the name people see: a signed record, attached to the file, of who made it and how it was edited. The signature makes tampering with the record detectable. Because the record is metadata, it can be lost when the metadata is removed, which often happens when you take a screenshot, re-save the file or run it through a tool like this one.

SynthID works the other way round. It carries very little information, but it lives in the pixels, so there is no record to delete. It is not signed and it holds no edit history. A detector has to look for it.

The two are complementary. Content Credentials can say a lot and are easy to lose. A pixel watermark says little and is harder to lose. A scan reads the first directly and can only infer the second.

How hidden watermarks are detected

A hidden watermark has two halves. The embedder nudges pixel values in a pattern that people cannot see. The detector knows that pattern and tests whether it is present. Without the detector, or the key behind it, the mark cannot be read.

Detectors usually return a confidence, not a plain yes or no. Edits such as heavy cropping, filters and re-compression lower that confidence, and a detector can be wrong in either direction.

We do not run these detectors. What we do is read the metadata. If it names a generator that is known to watermark its output, we say a hidden watermark is likely. That is an inference from the file, not a detection, and we label it that way in your report.

FAQ

Questions

What does the free scan show?

Three layers: visible marks in the pixels, AI metadata such as C2PA Content Credentials, and hidden pixel watermarks that we infer from the metadata. We cannot decode hidden watermarks, so a scan with no signals is not proof that an image is clean.

What is free?

The scan is always free. Quick clean is free up to 3 a day on the web, with output up to 12 MP. Original size above 12 MP, the GPU pass and the API use credits.

Will it remove SynthID?

The GPU pass redraws the image to disrupt hidden pixel watermarks such as SynthID. We cannot read these marks, so we infer them from the metadata, and we cannot promise a result. If it did not work for you, tell us on the result page: a GPU run gets one free re-run within 23 hours. If processing fails, your credits come back.

What do you keep in my file?

Creator and copyright fields stay in the file by default. We strip AI provenance fields only. If a Content Credential carries author values, the credential is removed and those values are carried into the standard fields.

What happens to my images?

Files are processed on our servers and on a GPU host we rent, not in your browser. Originals are deleted 24 hours after the last run on them finishes, or after upload if you never process (up to 72 hours while a job awaits payment). Results are deleted after 24 hours for anonymous visitors, 7 days for signed-in users and 72 hours for API jobs. You can delete an image and its results at any time with Delete now, and we never use images for training.

What if processing fails?

Your credits come back automatically, and a free-quota slot is returned. A run that finds nothing to remove is never charged.

Which files work?

PNG, JPG, JPEG, WEBP, AVIF. We scan images up to 100 MB (120 MB through the API) but only clean up to 25 MP. HEIC is not supported yet.

Is this allowed?

Use it on images you own or have the right to edit, and you confirm that before each clean. Some places, including the EU, China and India, require AI-generated or AI-edited content to be labelled in certain cases, and removing a mark from a file does not remove that duty. We refuse stock, marketplace and third-party marks and adult content, and the service is not available in mainland China. This is not legal advice.

Do you support video?

Not yet. It is images only for now.

Can I use the API?

Yes. Upload, scan for free, clean, then poll or receive a webhook when the job finishes. API cleans use credits and do not get the free daily Quick cleans. The reference is on the API page.

Does it work on edited images?

The scan reads what is in the file. Screenshots, crops and re-saves often drop metadata, so there may be less to read, and "no signals found" does not mean the image is clean. A visible mark can still be found if it is still in the picture. We cannot tell whether a hidden watermark survived an edit.

Can I undo a clean?

No. A clean is a new file, and the file you uploaded is never changed, so keep your own original. Our copy of the original is deleted 24 hours after the last run on it finishes.

Scan your image for free.

The scan costs nothing, and so do 3 Quick cleans a day. For images you own or have the right to edit.