AI Image Metadata Remover

An AI image can carry a surprising amount of information about how it was made: the prompt, the seed, the model and LoRA names, a complete ComfyUI workflow, a signed Content Credentials manifest, or an IPTC label that says “created by AI”. An AI image metadata remover finds those records, shows them to you, and deletes them without touching the pixels. Drop an image in the box above to see what yours holds. Files are cleaned in your browser and never uploaded.

Where AI information is stored in an image

Generators and editors do not agree on one place, so a single image can hold several kinds of records at once.

1. PNG text chunks. Free-form key and value pairs inside the PNG file. Most Stable Diffusion tools save the prompt and settings here.
2. EXIF, XMP and IPTC fields. Standard metadata blocks. They can hold a generator name, a description, an author, and a field that says the image is synthetic. JPEG and WebP outputs from many tools use these because those formats have no text chunks.
3. Content Credentials (C2PA). A signed manifest in its own container: a caBX chunk in PNG, APP11 segments in JPEG, a C2PA chunk in WebP.
4. JPEG comments and other segments. Small, easily forgotten blocks that some tools use for settings.
5. The pixels. Invisible watermarks such as SynthID. Not metadata, and not removable by a metadata tool.

What a Stable Diffusion PNG gives away

AUTOMATIC1111, Forge and similar web interfaces write one text chunk called parameters. It is plain text, and any viewer that reads PNG text can display it. An example, with made-up values:

a lighthouse at dusk, oil painting, warm light
Negative prompt: blurry, text, watermark
Steps: 28, Sampler: Euler a, CFG scale: 7, Seed: 1234567890, Size: 768x1024, Model hash: 1a2b3c4d5e, Model: sdxl-base-1.0, Lora hashes: "style-oil: 9f8e7d6c"

With the same prompt, seed and model, someone can regenerate your image or refine it. The model and LoRA names also say which checkpoints you use, which matters if they are private or licensed. When these tools save JPEG or WebP instead, the same text is usually written to an EXIF comment field, and it is removed there too.

What a ComfyUI workflow gives away

ComfyUI saves two text chunks in each PNG: prompt, the graph in execution form, and workflow, the full node layout as you see it on screen. Dragging such a PNG into ComfyUI restores the entire workflow, which is a feature when you want to share and a leak when you do not.

• Every prompt and seed typed into the graph.
• Checkpoint, LoRA, VAE and ControlNet file names.
• Names of input images and, in some setups, folder paths on your machine.
• The custom nodes you have installed, which shows your toolchain.
• A lot of text: a large workflow can add tens of kilobytes or more to a file.

Content Credentials (C2PA) in AI images

Content Credentials are built on the open C2PA standard. A manifest records which tool created or edited the file, which actions were taken, and who signed it, and it is bound to the image with a cryptographic hash. Several image generators and editors attach one by default, for example OpenAI’s image generation, Adobe Firefly and Photoshop’s generative features, and Microsoft’s design tools.

• It is a stored record, not a lock. Editing the pixels breaks the binding, and deleting the manifest leaves a normal image.
• Where it lives: a caBX chunk in PNG, APP11 segments in JPEG, a C2PA chunk in WebP, a box in MP4 and MOV.
• What it can name: the generating application, the signing organisation, timestamps, and the editing actions.
• A cloud copy may exist. Some providers keep their own copy and can look it up from an invisible watermark or a fingerprint. Deleting the manifest from your file does not change their records.

To see what a file claims, drop it on the Content Credentials verify page. After cleaning, the same page should report that none were found.

The IPTC digital source type

IPTC, the news-industry metadata body, defines a standard field called Digital Source Type. It is the simplest and most widely read AI label, and platforms look at it alongside Content Credentials. Three values matter here:

• trainedAlgorithmicMedia: the image was created by a generative model.
• compositeWithTrainedAlgorithmicMedia: part of the image was generated or edited with such a model.
• digitalCapture: the image came from a camera sensor.

The field lives in XMP. Cleaning removes the whole XMP packet, so the label goes with it, along with any creator-tool entry that names the generator.

How platforms use these signals

Meta, TikTok and LinkedIn have each said they read Content Credentials or IPTC fields to show labels on AI-made or AI-edited images, and YouTube asks creators to disclose realistic synthetic content when they upload. Two consequences follow. A label can be triggered by the metadata alone, even on a mostly real photograph that went through one AI-assisted edit. And metadata is only one input: platforms also rely on what you tell them and, in some cases, on detection. Cleaning a file does not remove an obligation to disclose when a platform asks you to.

What different sources typically leave behind

What cannot be removed

• Invisible watermarks, such as SynthID, which are part of the pixel values. No metadata cleaner removes them, and we do not claim to.
• Data hidden in the pixels. Some tools can encode their settings inside the picture itself. To a metadata tool it looks like an ordinary image.
• Visible marks: labels, logos and watermarks drawn on the picture.
• Provider records. The service you generated the image with still has your history and may hold its own copy of any credentials.

Why people remove AI metadata

• To keep prompts and workflows private. They are the work, and they can name clients and products.
• To avoid leaking your setup: model names, LoRAs, input file names and local folders.
• To shrink the file. Workflows and manifests add size you do not need on a website.
• To meet a client’s or a marketplace’s file requirements.
• To manage how an image is described. A real photo with one AI-assisted edit can inherit a label from its metadata. Editing the file does not change how you should describe the work.

Telling whether an image carries AI records

The quickest way is to drop the file in the tool above, because the list names every record it finds. A few hints help when you cannot:

• Stable Diffusion web interfaces include a “PNG info” tab. If it shows a prompt when you drop your image on it, the image holds text chunks.
• Images exported from OpenAI, Adobe or Microsoft tools should be assumed to carry Content Credentials.
• A PNG that is far larger than similar pictures may be carrying a long workflow. Compare its size with the cleaned copy.
• A Software, Creator or Description line in the file properties often names the generator.

Clean an AI image in three steps

1. Drop the image, or several. PNG, JPEG, WebP and AVIF work.
2. Press Clean and read the list. AI records are named plainly, for example “Generation parameters (prompt, seed, model)”, “ComfyUI workflow”, “C2PA Content Credentials”, “XMP packet” or “JPEG comment”.
3. Download. The pixels are identical to the original, and a second scan confirms nothing is left before the file is marked verified clean.

A whole generation run at once

A batch in ComfyUI or a web UI produces hundreds of PNGs, each with its own prompt and workflow. In the browser you can add a batch to the queue, up to your plan’s batch size, and download one .zip. For an output folder that keeps growing, the MCP server saves you the dragging: ask your assistant to clean the PNG files in the output folder, and each clean copy is saved next to its original as a new file. The API does the same from a script. Both leave the originals alone, so your prompts stay in the files you keep.

Keep your own record first

Once the records are gone from the file you shared, you cannot recover them from that file. Before you clean, save what you want to keep: store the prompt and settings in your own notes or project, export the ComfyUI workflow as a JSON file, and keep the original image in a private folder. Clean copies, not originals.

Check the result yourself

exiftool -a -G1 -s cleaned.png

Using it responsibly

Provenance records exist so that people can judge what they are looking at. Using this tool to protect your prompts and your privacy is fine. Using it to present generated images as photographs where authenticity matters, such as news, evidence, insurance or a contest that bans AI work, is not, and in a growing number of places and platforms AI-generated and AI-edited content must be disclosed whatever the file contains. Rules differ by country and change quickly, and this page is not legal advice. When your audience would expect to be told, tell them.

The AI metadata cleaner overview covers the wider picture, and the image metadata remover explains EXIF and GPS for ordinary photos.

Frequently asked questions