How it works

The records inside an image file

An image file is a container. Alongside the compressed pixels it carries blocks of descriptive data that no viewer shows you:

• C2PA content credentials — a cryptographically signed manifest describing how the file was made and edited, stored in a JPEG APP11/JUMBF segment, a PNG caBX chunk, a WebP C2PA chunk or an ISO-BMFF box in AVIF and HEIC.
• XMP — an XML packet holding the generator name, prompt, seed, sampler and edit history, including the digitalSourceType field whose value trainedAlgorithmicMedia is a direct machine-readable declaration that an image was AI-generated.
• EXIF — camera make and model, lens, exposure, timestamps, device serial numbers, GPS coordinates and a Software field that frequently names the AI tool.
• IPTC — credit, source and attribution fields Adobe and agencies use for AI disclosure.
• PNG text chunks — tEXt, iTXt and zTXt chunks where Automatic1111, Forge and ComfyUI store the entire generation workflow in plain text.

What the platforms actually read

Meta, TikTok and Pinterest all committed to labelling AI content, and the cheapest way to do that at scale is to read the records above at upload time rather than run a classifier over every image. That is why a photo can be labelled in under a second, and why an entirely real photograph gets labelled after one AI-assisted edit: the labeller sees the marker, not the picture. Creators hit this constantly with Photoshop Generative Fill, Lightroom AI Denoise, AI background removal and CapCut exports.

Why re-encoding removes everything

AI Label Remover does not try to find and delete individual records — that approach misses whatever it does not know about. Instead the image is decoded into raw pixels with the browser’s own image decoder, and those pixels are encoded into a brand-new file. Metadata is not part of the pixel data, so nothing survives the round trip. One pass removes C2PA, XMP, EXIF, IPTC, PNG chunks, JPEG comments and anything else the container was carrying, including formats that did not exist when this page was written.

The four steps

1. Drop your image. The file is read into browser memory with the File API. No network request is made — you can watch the network tab stay empty.
2. Inspect what is inside. The file’s container is parsed on the spot to list every C2PA, XMP, EXIF, IPTC and PNG-chunk record it carries, and to name the generator that wrote them.
3. Decode and re-encode. The picture is decoded to raw pixels and encoded into a brand-new file. Metadata cannot survive that round trip, so every hidden record is gone in one pass.
4. Download and post. You get a clean file with identical dimensions, optionally with a reset fingerprint and placeholder camera EXIF, ready to upload anywhere.

Fingerprint reset

Platforms also compute a perceptual hash of an upload so they can match it against copies of the same image seen elsewhere — which means a previously labelled file can be recognised even after cleaning. The optional fingerprint reset shifts a scattered subset of pixels by ±1 RGB. The change is invisible at any zoom level, but it is enough to move the hash outside the match threshold.

Placeholder camera EXIF

A file with no metadata at all is itself slightly unusual, since ordinary camera exports carry a basic EXIF block. The optional injection writes a plain, non-identifying block — make, model, software version, timestamp and orientation — after cleaning, so the export looks like a normal photo export. It contains no GPS, no serial number and nothing traceable to you. JPEG output only.

Privacy

There is no upload endpoint in this application. Files are read with the File API, drawn on a canvas and encoded with canvas.toBlob, all inside your browser tab. You can open your network panel and watch it stay empty, or disconnect from the internet after the page loads and clean images offline. Read the privacy policy.