Private browser tool · No signup · No upload

AI Metadata Remover

See exactly what an AI image says about how it was made: the prompt, the workflow, the generator and its Content Credentials. Then remove that metadata without touching a pixel.

Updated

Drop AI images here

JPG, PNG and WebP are cleaned without re-encoding. HEIC is converted to JPG.

Processed in your browser. Nothing is uploaded.

An AI metadata remover deletes the information a generator writes into the file: prompts, settings, workflows, generator names and Content Credentials. Drop the image above, read the found card, and click Remove all metadata. The pixels are copied unchanged, the cleaned file is checked again, and nothing is uploaded.

There are ordinary reasons to want this. A Stable Diffusion PNG can carry your full prompt, negative prompt and seed; a ComfyUI PNG carries the entire node graph you spent an evening building. Client deliverables are cleaner without either. And as an AI image metadata remover, the found card is useful by itself: drop an image someone sent you to see what it says about how it was made, and change nothing.

What this does not do matters as much, and the section on what stays is plain about it: an image without metadata can still be identified as AI-made.

What AI images carry in their metadata

Generators label their output in four different places, and an image can carry several at once. The AI metadata remover looks in all four.

Where AI information lives in the file

WhereWhat it looks likeWritten by
Content Credentials (C2PA)A signed manifest: JPEG APP11 segments, a PNG caBX chunk, a WebP C2PA chunkChatGPT and the OpenAI API, Adobe Firefly, Microsoft Designer
IPTC Digital Source Type (in XMP)trainedAlgorithmicMedia or compositeWithTrainedAlgorithmicMediaTools following the IPTC vocabulary
PNG text chunksparameters (Automatic1111, Forge); prompt and workflow JSON (ComfyUI); Comment and Software (NovelAI)Local Stable Diffusion tools
EXIF and XMP text fieldsSoftware, ImageDescription, CreatorTool or description naming a generatorVarious

Sources: C2PA specification 2.1, IPTC Digital Source Type NewsCodes, checked November 2026.

The IPTC vocabulary draws a line worth knowing. trainedAlgorithmicMedia means media created by a model trained on sampled content; compositeWithTrainedAlgorithmicMedia means a real image altered with generative AI, such as by inpainting or outpainting. A photo with one generative-fill patch can carry the second label.

Content Credentials are a signed record of who or what made the file. The C2PA specification has the signing tool identify itself in a claim_generator string, or in newer 2.x manifests the name inside claim_generator_info. The card reads both, and when they name a product, it shows it. This is a best-effort read of the manifest, not a signature check.

People look for a ChatGPT image metadata remover or a Content Credentials remover; this AI metadata remover is both, and it reads local Stable Diffusion files too.

AN AI-GENERATED PNG · CHUNK BY CHUNK IHDRWidth, height, colour typekeptiCCP · sRGB · gAMAColour informationkepttEXt "parameters"Prompt, negative prompt, seed, modelremovedtEXt "prompt" · "workflow"ComfyUI graph as JSONremovediTXt XMPDigitalSourceType, CreatorToolremovedcaBXContent Credentials (C2PA)removedIDATThe pixelsuntouched Anything hidden inside the pixels themselves, such as a watermark, is in IDAT and is left exactly as it was.
An AI-generated PNG chunk by chunk. Orange rows are removed; colour chunks and pixels stay.

Check what the file says, then strip it

For a PNG with prompt chunks, the found card shows the prompt text itself, not just "prompt found", so you see exactly what anyone receiving the file could read. For ComfyUI, the workflow JSON is what lets someone drag your PNG onto their canvas and rebuild your node graph; the ComfyUI SaveImage documentation (checked November 2026) describes the node embedding workflow metadata in the saved file. Automatic1111 saves its parameters chunk by default; its settings call this "Save text information about generation parameters as chunks to png files".

  1. Keep the original. The tool never changes your file; it creates a cleaned copy. Keep the original if you want to reproduce the image later, because the prompt and seed are gone from the copy.
  2. Read the card before you click. Sometimes the only thing in the file is a colour profile, and there is nothing to remove.
  3. Strip and look at the re-check. The cleaned file is parsed again from scratch; "Nothing left" is a reading, not a promise.
  4. Download one file or a ZIP. A batch keeps each file's own name.

JPG, PNG and WebP all lose every item in the table above, Content Credentials included. HEIC files are read for the card first, then converted to JPG.

WHAT THE FOUND CARD SHOWS · NOTHING CHANGED YET Content Credentialspresent · names a generatorDigitalSourceTypetrainedAlgorithmicMediaPNG text: parametersshown belowSoftwarenames a generator a lighthouse at dusk, oil painting, warm light Negative prompt: blurry · Steps: 30 · Seed: 1234 …exactly the text anyone you send the file to can read
The AI found card works as a viewer: the prompt is shown as text before anything is removed.

What an AI metadata remover cannot remove

Metadata is one layer of evidence among several, and the easiest to lose. OpenAI's Help Center (checked November 2026) says C2PA metadata is easy to remove, by accident or on purpose, and that images from ChatGPT and its API therefore carry a SynthID watermark as well.

Invisible watermarks stay. Google DeepMind (checked November 2026) describes SynthID as an invisible watermark designed to withstand cropping, filters and lossy compression, and says its SynthID Detector checks content from Google and partners including OpenAI. The watermark is in the pixels; this tool copies pixels unchanged, so the watermark is exactly where it was.

Pixel-level data stays. NovelAI also writes generation settings into the alpha channel of its PNGs, according to NovelAI's own metadata repository (checked November 2026). That data is part of the image, not a chunk; this tool neither detects nor removes it.

The image stays recognisable. Detectors that analyse the picture itself work without any metadata. Say it plainly: a cleaned image can still be identified as AI-made.

WHAT CAN SHOW AN IMAGE IS AI-MADE File metadataC2PA manifest, IPTC source type, prompts, generator namesThis tool removes itInvisible watermarkSynthID and similar marks woven into the pixelsStaysThe image itselfVisual traits any detector can analyseStaysYour disclosurePlatform rules to label realistic AI contentStill yours
Four layers that can show an image is AI-made. Only the first is file metadata.

Platform rules do not change

The big platforms read Content Credentials to label AI content automatically, but none of them relies only on metadata, and all of them ask creators to disclose realistic AI content themselves.

AI labelling at three platforms

PlatformReads metadataWhat it asks of you
Facebook, Instagram, ThreadsC2PA and IPTC signalsUse the AI label for photorealistic video and realistic audio; penalties possible
YouTubeC2PA; such labels can't be changedDisclose realistic altered or synthetic content with the AI setting at upload
TikTokC2PA Content CredentialsLabel realistic AI-generated content

Sources: Meta Newsroom (updated April 2025) and Instagram Help Center; YouTube Help; TikTok Newsroom. All checked November 2026.

Using an AI metadata remover changes none of this. If you post a realistic AI image or video where a platform asks for disclosure, label it yourself. YouTube, for example, asks for disclosure when content makes a real person appear to do something they didn't, alters footage of a real event, or shows a realistic scene that never happened, and exempts clearly unrealistic content and minor aesthetic edits.

AI metadata remover — questions

Does this remove the Content Credentials from ChatGPT images?

It removes the C2PA manifest from JPG, PNG and WebP files, which is where Content Credentials live. OpenAI says its images also carry a SynthID watermark in the pixels; that stays.

What about Adobe Firefly images?

Adobe states that Content Credentials are applied automatically to content generated with Firefly. The card shows the manifest and, when it names one, the generator; removal drops the manifest from the file.

Does Microsoft Designer add metadata?

Microsoft says images generated with AI features in Designer carry content credentials, and its download options include Manage content credentials. The card reports them as Content Credentials (C2PA) present.

Will it remove my Stable Diffusion prompt?

Yes. Automatic1111 and Forge store the prompt, negative prompt and settings in a PNG text chunk called parameters. The card shows that text, and removal drops the chunk. Keep your original if you want to reproduce the image.

Does it remove a ComfyUI workflow?

Yes. ComfyUI saves prompt and workflow JSON in PNG text chunks. Both are shown on the card and both are removed, so the cleaned PNG no longer loads as a workflow.

What about NovelAI images?

The Comment and Software text chunks are removed. NovelAI can also store settings in the alpha channel of the pixels; that is image data, so this tool leaves it untouched and does not report it.

Can someone still tell my image is AI-generated?

Yes. Invisible watermarks such as SynthID stay in the pixels, and detectors can analyse the image itself. Removing metadata removes one label, not the evidence, and platform labelling duties stay the same.

Can I remove AI metadata from a photo on my phone?

Yes. The AI metadata remover runs in current mobile browsers on iPhone and Android. Pick images from your photo library, strip them, and save the cleaned copies or a ZIP.

Are my images uploaded?

No. Everything is read and rewritten in your browser. Analytics count anonymous events, such as which kind of AI metadata was found, but never image data, prompts or filenames.