How to remove AI metadata from an image

Ask an AI tool for an image and you usually get more than pixels. The file can name the model that made it, state in a standard vocabulary that it was generated by a trained algorithm, record who signed that statement, and, depending on the tool, carry the exact prompt, seed and settings you used. None of it shows on screen. All of it travels with the file when you send it.

This guide takes each kind in turn: what it says, where it sits, how to see it and how to take it out cleanly, plus what removal cannot do.

What counts as AI metadata in an image?

Four separate mechanisms are in common use. They come from different standards and different tools, so one image can carry all four, any mix of them, or none at all.

• C2PA Content Credentials. A signed manifest naming the app or service that produced the file, listing what was done to it and, where it applies, declaring it AI-generated. Adobe Firefly, OpenAI’s image tools and Microsoft Designer attach one. A JPEG keeps it in APP11 segments, a PNG in a caBX chunk, a WebP in a C2PA chunk, and HEIC or AVIF in a uuid box.
• IPTC labels. The IPTC digital source type is a short value, usually stored in the image’s XMP packet: trainedAlgorithmicMedia means made by a generative model, compositeWithTrainedAlgorithmicMedia means a real photo with AI-generated parts. Newer IPTC fields go further and record the AI system used, its version and the prompt.
• Generator tags. Ordinary fields that name the software that wrote the file, such as EXIF Software, XMP CreatorTool, a PNG text entry or a JPEG comment, filled in with a generator’s or AI editor’s name.
• Prompts and settings. Stable Diffusion front ends save the recipe. AUTOMATIC1111 writes the prompt, negative prompt, seed, sampler and model hash into a PNG text chunk called parameters, and ComfyUI stores its entire node graph in prompt and workflow chunks.

The first two are deliberate disclosure systems: the image declaring what it is. The last two are side effects of how software saves files, which is why they tend to reveal more than anyone intended. A prompt can contain a client’s name, a real person’s name or wording you would never publish, and a ComfyUI workflow lists the model files and input images it loaded from your machine.

Why would you remove AI metadata?

• Privacy. A manifest can name the account or app that signed it and include thumbnails of the source images, and prompt chunks show exactly what you typed.
• A real photo, mislabeled. A small generative fix, such as removing a stray object or extending a background, is enough for some editors to record an AI edit, and platforms that read that record may label the whole photo as AI.
• Clean deliverables. Agencies and freelancers often hand over files without the tool chain and prompt history of the job embedded in them.

How do you check an image for AI metadata?

The file viewers built into Windows and macOS show basic EXIF fields, but they were not built to decode C2PA manifests or the text chunks Stable Diffusion tools write. Use a reader that understands all four signals:

1. 1Open the metadata viewer, or the AI metadata remover, which runs the same inspection, and drop the image in. It is read in your browser, not uploaded.
2. 2Look at the AI & provenance group first. When a manifest is present, a Content Credentials card shows who signed it, the app that made the file and the recorded actions, with an AI-generated badge if an action declares an AI source type.
3. 3Read the rows under it: digital source type, AI system and prompt fields, generator tags and prompt or workflow chunks. Each carries a badge saying where it was found, such as EXIF, XMP, IPTC or PNG text.
4. 4Check the other groups before you pick a mode. If Location lists GPS coordinates, AI-only removal will keep them.

An empty AI & provenance group means the file declares nothing about AI. It does not mean the image is a real photo: screenshots, many re-saved files and most images downloaded from social networks arrive with little or no metadata at all.

Which removal mode should you use?

Toolvy’s metadata tools have two modes, and the choice matters most for photos, which carry plenty of data that has nothing to do with AI. AI & provenance only edits the XMP and IPTC blocks surgically, taking out the AI properties and leaving the rest exactly as written. Everything removes every metadata field.

Choose AI-only when it is your own photograph and you want to keep the camera record and your copyright line. Choose Everything when the file is going somewhere public and nothing in it needs keeping. Its one visible side effect: an Ultra HDR JPEG loses its gain map and shows in standard range on an HDR screen.

How do you remove it step by step?

1. 1Open the AI metadata remover and drop one image or a whole batch: JPEG, PNG, WebP, GIF, AVIF and HEIC are all accepted.
2. 2Leave the mode on AI & provenance only, or switch to Everything if location and camera data should go as well.
3. 3Press the remove button. Each file is rewritten at the byte level, so nothing is re-encoded.
4. 4Read the result. Every cleaned copy is inspected again by the same reader, so you see what is left rather than a promise. The AI & provenance group should now be empty.
5. 5Download the files one at a time or as a ZIP. Copy report puts a plain-text summary of what was found and removed on your clipboard.

If all you want gone is the signed manifest, the C2PA remover opens in the same AI-only mode with the emphasis on reading the Content Credentials first. For an ordinary photo with GPS and camera data and no AI involved, the EXIF remover is the quicker route.

What can removal not do?

• Invisible watermarks stay. Google’s SynthID and similar systems hide a signal in the pixels themselves. That is not metadata, so it survives cleaning. Toolvy does not alter pixels and does not try to remove watermarks.
• Online credentials stay. Some services keep a copy of the Content Credentials on a server and can find it again from an invisible watermark or a fingerprint of the image, which C2PA calls a soft binding.
• Platforms can still detect AI. Besides reading metadata, large platforms run their own classifiers, so a fully generated image may be labeled regardless.
• Shared copies are unchanged. Cleaning produces a new file. The original, and any copy you already sent, still carries everything.

A screenshot or a re-save also drops most metadata, but it re-encodes the picture and leaves no record of what was removed. A byte-level clean keeps every pixel and tells you exactly what changed.

Frequently asked questions

Does removing AI metadata reduce image quality?

No. The metadata blocks are cut out of the file and the image data is copied unchanged, so every pixel is bit-for-bit identical. The file usually gets slightly smaller.

Will my image still be labeled as AI on Instagram?

If the label came only from metadata, a cleaned copy uploaded fresh has nothing for that check to read. But Meta also looks for invisible watermarks and uses its own detection, so a fully generated image can still be labeled, and posts already published do not change.

How do I remove the prompt from a Stable Diffusion PNG?

Drop the PNG into the AI metadata remover. The parameters chunk from AUTOMATIC1111, the prompt and workflow chunks from ComfyUI and InvokeAI’s metadata chunks are listed under AI & provenance and removed in either mode, with the pixel data left untouched.

Is my image uploaded?

No. Inspection, removal and the second check all run in your browser tab; DevTools → Network shows no request carrying your file.