Image Metadata Remover — EXIF, C2PA & AI Label Stripper

Bulk image metadata remover for Apify. Remove metadata from photos and images — EXIF, GPS, IPTC, XMP, ICC, Stable Diffusion prompt chunks, C2PA Content Credentials and the "generated by AI" label — in one run.

The default mode is lossless: the decoded pixels of the output are bit-identical to the input (verified, mean delta 0). Only the metadata containers are removed.

What metadata does it remove?

Layer	Examples	Handled
EXIF	GPS coordinates, camera Make/Model, serial numbers, timestamps, Software	Yes
IPTC	Caption, credit, byline, copyright	Yes
XMP	Creator, rights, editing history, DigitalSourceType (the AI label)	Yes
PNG text chunks	Stable Diffusion / ComfyUI / A1111 prompt, seed, model, sampler	Yes
WebP / HEIC / AVIF boxes	EXIF and XMP RIFF/ISOBMFF chunks	Yes
C2PA / JUMBF	Content Credentials manifests and signatures (APP11 / jumb boxes)	Yes
ICC profile	Optional — keep it with keepColorProfile	Yes
Embedded thumbnails	The small preview that often survives naive strippers	Yes
Classic invisible watermarks	DWT-DCT style frequency-domain marks	scrub mode
SynthID, Stable Signature, Tree-Ring	Model-space watermarks	No — flagged, not removed

Supports JPEG, PNG, WebP, HEIC, AVIF, TIFF and GIF.

How to remove metadata from photos in bulk

Send a list of image URLs and pick a mode. One dataset record per image is returned, plus the cleaned file in the run's key-value store.

{
  "imageUrls": ["https://example.com/photo.jpg"],
  "mode": "clean",
  "outputFormat": "auto",
  "concurrency": 5
}

Images can also come in as imagesBase64, or as keyValueStoreKeys from an existing key-value store (sourceKeyValueStoreId).

Three modes:

• inspect — report only. Lists every layer, every tag, whether a C2PA manifest or an AI label is present. The file is not modified.
• clean (default) — remove all metadata and C2PA. Pixels untouched.
• scrub — clean, plus a pixel-level pass (re-encode, resize round-trip, light noise, optional border crop) that destroys classic invisible watermarks.

Remove the AI-generated label and C2PA Content Credentials

Modern AI generators tag their output. The IPTC/XMP DigitalSourceType = trainedAlgorithmicMedia field marks an image as AI-made, and C2PA Content Credentials embed a signed provenance manifest (JUMBF boxes / APP11 segments) that can name the generator, the account and the full edit history.

This Actor strips both: the AI label tags and the entire C2PA manifest. The manifest is bound to the exact byte sequence of the file, so dropping it invalidates the signature and removes the provenance chain outright — no C2PA library needed, no pixel touched.

For classic invisible watermarks (DWT-DCT frequency-domain marks), scrub mode applies a pixel-level attack. Verified against a real DWT-DCT watermark: decode accuracy falls from 100 % to 42 %, where random guessing sits at 50 % — nothing recoverable. For scrub runs, set outputFormat to webp or jpeg; PNG cannot compress the added noise and the file grows a lot.

Does it change my pixels?

No, not in the default clean mode. The decoded image is bit-identical before and after (mean/max pixel delta 0 across the test set); only the metadata containers are dropped.

scrub mode intentionally changes pixels — that is how it breaks invisible watermarks. Every scrub record reports the exact pixel mean/max delta and byte sizes so you can see how much the image changed.

Output

One dataset record per image with the layers found before and after, the full tag dump before and after, aiLabelRemoved, c2paRemoved, clean, pixelChanged, pixel mean/max delta, byte sizes, warnings, and a public link to the cleaned file in the run's key-value store.

Limitations

• Google SynthID, Meta's Stable Signature and Tree-Ring watermarks are embedded in the generative model itself, not in the file. There is no cheap CPU method to remove them, so this Actor does not pretend to. When the source metadata points at such a generator, the record carries an explicit warning instead.
• clean deliberately does not touch pixels, so any pixel-domain watermark survives it — use scrub for those.

FAQ

Does it reduce image quality? No. The default clean mode is lossless — pixels are bit-identical. Scrub mode changes pixels on purpose and reports the exact difference.

Can it remove metadata from AI-generated images? Yes. It removes the IPTC/XMP AI label (DigitalSourceType), Stable Diffusion / ComfyUI prompt chunks, and C2PA Content Credentials left by generators such as Adobe Firefly.

Can it remove the Gemini / SynthID watermark? No — SynthID lives in the model, not the file, and no cheap removal exists. The Actor flags a warning when an image likely carries it instead of claiming a fake removal.

What does it cost? Fractions of a cent per image on the Apify platform — roughly $0.00002 per image at 1 GB RAM, plus any Actor pricing if set.

Full measurements: docs/ai-footprint-remover-poc.md.

For AI agents (MCP-ready)

This actor is built to be called by AI agents. It works out of the box with the Apify MCP Server — add it to Claude Desktop, Cursor or any MCP client: the agent can strip image metadata and AI provenance labels on its own.

{
  "mcpServers": {
    "ai-footprint-remover": {
      "url": "https://mcp.apify.com/?actors=andrew_babo/ai-footprint-remover",
      "headers": { "Authorization": "Bearer <YOUR_APIFY_TOKEN>" }
    }
  }
}

Agent skill (paste into your agent's instructions)

Use the "ai-footprint-remover" tool to strip EXIF/XMP/IPTC metadata, C2PA
provenance manifests and AI-generation labels from images, and optionally
disturb invisible watermarks. Typical asks: "remove metadata / GPS from
these photos", "clean this AI image before publishing".

HOW TO CALL
- { "imageUrls": ["https://.../a.jpg"], "mode": "clean" }
- "mode": "clean" strips metadata only (lossless-ish, default);
  "deep" additionally re-encodes, resizes, adds light noise and crops a few
  pixels to disturb invisible watermarks — use only when the user asks.
- Batch up to a few dozen images per run; "concurrency" defaults to 5.
- Keep "storeOutput": true to get downloadable cleaned files.

OUTPUT CONTRACT
- One row per image: output URL, removed metadata keys, detected C2PA /
  AI labels, before/after size.
- errorCode on a row means that image failed (404, too large, unsupported
  format) — report which ones failed instead of claiming all were cleaned.
- Never promise that invisible watermark removal is guaranteed; "deep" mode
  degrades watermarks, it does not certify their absence.