Free · no upload · no account
Remove the AI label from your photos, without touching a pixel.
Most tools tell you a file is clean without reading it, then re-encode your image anyway. This one shows you every C2PA, XMP, EXIF and PNG field it finds, strips what you choose, and proves your pixels came through byte-for-byte identical.
- 1Drop in a JPEG, PNG or WebP
- 2See every marker we find, with its real values
- 3Download a clean copy, same pixels, smaller file
Drop an image here
We read what is inside it first, then show you every field before anything is removed.
JPEG · PNG · WebP: up to 25.0 MB · 30 files at a time
Nothing is uploaded
Parsing and stripping run in a Web Worker on your machine. Switch off your network and it still works.
Nothing is re-encoded
Metadata is removed by container surgery. The compressed pixel data is copied, never decoded.
Nothing is assumed
Every result is re-scanned afterwards. You only see “verified clean” when the check actually passed.
What is actually in there
The five things hiding in your image file
An image is not only pixels. Alongside the compressed picture, most files carry several separate metadata containers, written by different software for different reasons. Tools that only handle one of them will call your file clean while the other four sit untouched.
- EXIFExchangeable Image File Format
- Written by cameras, phones and most editors. Carries make and model, lens, shutter and aperture, the exact date and time, and, on a phone: GPS coordinates accurate to a few metres. It also often carries a Software field, which is where editors quietly announce themselves: "Adobe Photoshop 25.0 (Generative Fill)" is EXIF, and it is enough to get a real photograph labelled.
- XMPExtensible Metadata Platform
- Adobe's XML block, and the usual home for generation parameters. Midjourney writes prompt text and seed here. Adobe writes IPTC DigitalSourceType, whose value trainedAlgorithmicMedia is a machine-readable declaration that the image was AI-generated. Platforms read this field directly.
- C2PACoalition for Content Provenance and Authenticity
- A cryptographically signed manifest describing how the file was made and edited. OpenAI, Adobe, Microsoft, Google and a growing list of camera makers write it. In JPEG it lives in an APP11 JUMBF box; in PNG, a caBX chunk. This is the single marker most likely to trigger an AI label on upload.
- IPTCInternational Press Telecommunications Council
- Publishing fields: creator, credit, copyright, caption, usage terms. Legitimate and often worth keeping, which is why our AI-only mode leaves camera and attribution data alone. Some AI tools also write AI source-type values into IPTC.
- PNG texttEXt / zTXt / iTXt chunks
- Stable Diffusion and ComfyUI store entire workflows in plain text chunks inside the PNG, prompt, negative prompt, sampler, CFG scale, seed, model hash, and in ComfyUI's case the full node graph as JSON. Anyone who opens the file in a text editor can read your prompt.
Where it comes from
Which AI tools leave which markers
Different generators write to different places, which is why a tool that checks only one container misses so much. This is what we see most often:
| Tool | Writes to | What it leaves |
|---|---|---|
| Midjourney | XMP | Prompt text, seed, version flags |
| ChatGPT / DALL·E | C2PA | Signed OpenAI content credentials |
| Adobe Firefly | C2PA + IPTC | Content credentials and AI source type |
| Stable Diffusion | PNG text | Full generation parameters in tEXt |
| ComfyUI | PNG text | Entire node workflow as JSON |
| Photoshop Generative Fill | C2PA + EXIF | Partial-edit credentials on real photos |
| Lightroom AI Denoise | EXIF + XMP | Software string and edit history |
| Google Gemini / Imagen | C2PA | Content credentials, plus SynthID in pixels |
| Leonardo AI | XMP | Generation signatures |
| Ideogram | EXIF | Generator markers in EXIF |
| Canva AI | XMP + C2PA | Varies by export path |
| Runway | C2PA | Signed credentials on stills and frames |
Why you got labelled
What each platform reads on upload
Platforms do not usually judge your pixels. They read the provenance markers in the file you handed them, which is why a photograph you took yourself can be labelled after one AI-assisted edit.
- Instagram & Facebook
- “AI info” label
Reads C2PA and XMP on upload. The most aggressive of the major platforms.
- TikTok
- “Contains AI-generated media”
Checks cover images, Photo mode stills and carousel slides.
- AI-modified label
Reads C2PA; commonly affects product and mockup pins.
- C2PA provenance notice
Shows content credentials when present.
- X / Twitter
- Varies
Inconsistent, but C2PA is read on some upload paths.
The part nobody else gets right
Lossless removal, and how to check it
Metadata sits in length-prefixed blocks beside the compressed pixel data, APP segments in JPEG, chunks in PNG and WebP. Removing it is a matter of copying the file forward and skipping those blocks. No decoder, no encoder, no generation loss.
Most tools instead decode the image into a canvas and re-save it. That strips metadata as a side effect, and it is much less work to build, but it also re-compresses your photograph. We measured a widely used competitor on a single pass:
| pixels altered | 52.3% | 0% |
| PSNR | 39.78 dB | ∞ (identical) |
| max channel difference | 46 / 255 | 0 |
| PNG output size | +15.8% | smaller than input |
left column: a leading competitor · right column: this tool
You do not have to take that on trust. After every strip we parse the output again and compare the pixel payload byte-for-byte against the input. The “verified clean” line only appears when that comparison, the size check and the zero-remaining-fields check all pass.
Who this is for
Six reasons people end up here
Photographers
You used AI Denoise or a background fill on a real photograph, and the deliverable now carries edit provenance. AI-only mode strips the generator markers and leaves your camera settings, creator and copyright fields intact.
Etsy and print-on-demand sellers
Listing images get flagged, or carry the GPS of the room you photographed them in. Batch-clean a folder of mockups before upload.
Social media managers
Exports from Canva, CapCut and Photoshop carry provenance you did not add. Check a sample before scheduling a whole campaign.
Journalists and researchers
GPS, device serial numbers and timestamps identify where you were and what equipment you used. This is the original reason EXIF stripping exists.
AI artists
Your prompt and seed are sitting in the file in plain text. If you sell or publish the image, you are publishing your method with it.
Developers
The engine is a dependency-free TypeScript module that runs unchanged in Node and the browser. Read the source rather than trusting the claim.
False positives
When a real photograph gets called AI
This is the most common reason people arrive here, and the most frustrating. A wedding photographer runs Generative Fill to extend a background. A product shot gets an AI sky replacement. A portrait goes through Lightroom AI Denoise. A Reel cover is trimmed in CapCut. In each case the photograph is real, and in each case the editor wrote provenance into the file saying an AI tool touched it.
Platforms read that marker and apply a label. They are not wrong about the metadata: the edit did happen, but the label implies something about the image that is not true. Removing the marker from a file you own puts the file back to describing what it actually is.
If the checker finds nothing and the label still appears, the cause is not metadata. At that point you are looking at a pixel-level classifier or an invisible watermark, and no metadata tool, this one included: will change it.
Worth knowing
Legal and compliance, honestly
Stripping metadata from your own files is ordinarily lawful, and removing GPS and device identifiers before publishing is routine privacy hygiene, it is why EXIF removers have existed for twenty years.
Disclosure is a different question from metadata. The EU AI Act places transparency obligations on providers and deployers of generative systems, and several platforms require AI disclosure in their own terms. Removing a marker does not remove an obligation that applies to you. If you generated an image with AI and the context requires you to say so, say so.
Where this tool is unambiguously useful: reclaiming a real photograph that was mislabelled, and stripping location and device data before you publish. None of this is legal advice.
Limits
What this deliberately cannot do
- It never touches pixels. Invisible watermarks like SynthID survive, because they live in the image data itself. That is a consequence of the lossless design, not an oversight.
- It does not validate C2PA signatures. We detect and remove the manifest, but do not yet verify who signed it.
- AVIF and HEIC are not supported yet. They need ISOBMFF box handling. We would rather reject a format than silently convert it to PNG, which is what some tools do.
- Compressed PNG text is removed but not displayed. zTXt and iTXt chunks are stripped; we do not yet inflate them to show contents.
Questions
Frequently asked questions
›Does this upload my images?
No. Parsing and stripping both run in your browser, in a Web Worker. No image bytes are sent anywhere, there is no server endpoint that could receive them. You can verify this: turn off your network connection and the tool keeps working, because it never needed the network in the first place.
›Will removing metadata reduce my image quality?
No, and this is the main thing that separates this tool from most others. Metadata lives in length-prefixed blocks that sit beside the compressed pixel data, so removing it means copying the file forward and skipping those blocks. The compressed image data is never decoded or re-encoded. We re-check the output after every strip and show you a "pixels unchanged" confirmation.
›Why do other tools say "same pixels" when they re-encode?
Because decoding and re-saving is far easier to build. We measured one popular tool: it altered 52.3% of pixels, with a peak signal-to-noise ratio of 39.78 dB and a maximum per-channel difference of 46 out of 255, while its homepage claimed "same pixels". Its PNG output also came back 15.8% larger than the input. That is a generation of loss added for no reason.
›What is the difference between "Everything" and "AI markers only"?
"Everything" removes every removable block: EXIF, GPS, XMP, C2PA, IPTC and PNG text. "AI markers only" removes C2PA, XMP and PNG text, and removes EXIF only when it carries a generator fingerprint, so a photographer keeps shutter speed, lens and copyright. Note that AI-only mode keeps GPS, because GPS lives in EXIF. If privacy is your goal rather than labelling, choose Everything.
›Will this stop Instagram showing the AI label?
It removes the file-level markers that most commonly trigger it, C2PA and XMP. That handles the large majority of cases, including real photographs wrongly flagged after an AI-assisted edit. It cannot help if a platform is using pixel-level analysis or an invisible watermark, because those are not metadata and no metadata tool touches them.
›Can you remove SynthID or other invisible watermarks?
No, and you should be sceptical of any tool that says it can. SynthID and similar schemes embed a signal in the pixels themselves. This tool deliberately never modifies pixels, that is the whole design. If an image is watermarked at the pixel level, stripping metadata will not change that.
›Is it legal to remove metadata from my own images?
Removing metadata from files you own is ordinarily lawful, and stripping GPS and device identifiers before sharing is standard privacy practice. Disclosure rules are a separate question: the EU AI Act places transparency obligations on providers and deployers of generative systems, and some platforms require disclosure in their terms. Removing a marker does not remove an obligation to disclose where one applies. We are not your lawyers.
›What file formats are supported?
JPEG, PNG and WebP, up to 25 MB per file and 30 files at a time. AVIF and HEIC are not supported yet, they use the ISOBMFF box structure, which needs its own implementation. We would rather not support a format than convert it to something else behind your back.
›Why does my file get smaller but not much smaller?
Because metadata is usually a tiny fraction of the file. A typical EXIF block is a few hundred bytes to a few kilobytes against several megabytes of pixel data. Seeing a small reduction is the expected, correct result, a large reduction would mean something re-compressed your image.
›What does the verification receipt actually check?
After stripping, the output is parsed again with the same engine. We confirm the pixel payload is byte-for-byte identical to the input, that the output is no larger than the input, and that zero removable metadata fields remain. All three have to hold before you see "verified clean".
›Can I keep my copyright and attribution?
Yes: choose "AI markers only". Creator, credit and copyright fields in EXIF and IPTC are preserved unless they themselves contain a generator fingerprint.
›Does it work offline?
Yes. After your first visit the page and its code are cached, and because the engine makes no network calls, everything keeps working with the network off. This is the most direct way to confirm that nothing is being uploaded.
›Do you detect C2PA signatures as valid or invalid?
Not yet. We detect that a C2PA manifest is present and remove it, but we do not currently verify its cryptographic signature or show you who signed it. Signature validation is planned. We would rather state this plainly than imply verification we do not perform.
›Why does a photo I took myself get an AI label?
Almost always because an AI-assisted edit left provenance behind. Generative Fill, AI sky replacement, AI background removal, Lightroom AI Denoise and several CapCut effects all write C2PA or XMP into an otherwise ordinary photograph. The pixels are yours; the marker came from the editor.
›Is there a limit on how many images I can clean?
Thirty files per batch, and no limit on batches. There is no account, no sign-up and no quota, because there is no server doing the work, your own machine is doing it.
›How is this different from an ordinary EXIF remover?
An EXIF remover handles one of the five things in a modern image file. C2PA manifests, XMP generation blocks and PNG text chunks are not EXIF, and tools that only handle EXIF will leave all of them in place while reporting the file as clean.