How to Remove C2PA Metadata from DALL-E and Adobe Firefly Images

If you have ever opened a DALL-E 3 image in a tool like ExifTool or Adobe Bridge, you may have noticed a block of data labeled 'C2PA' or 'Content Credentials' that looks nothing like ordinary EXIF fields. This is the Coalition for Content Provenance and Authenticity standard — a cryptographically signed provenance system that has become the most technically sophisticated form of AI metadata in widespread use. In 2026, C2PA credentials are embedded by default in images from DALL-E 3, Adobe Firefly, Bing Image Creator, and an expanding list of other AI tools.

This guide is written for creators, developers, and privacy-conscious users who need to understand exactly what C2PA is, what information it contains, and how to remove it completely from their images. We will cover the technical structure of C2PA data, why standard EXIF stripping tools fail to remove it, and the correct approach to complete credential removal.

Understanding the C2PA Technical Structure

C2PA data is stored in a JUMBF (JPEG Universal Metadata Box Format) container — a binary structure that is embedded in JPEG files as an APP11 marker segment and in PNG files as a dedicated chunk. Unlike EXIF data, which is stored as a relatively flat key-value structure, C2PA data is a nested JSON-like manifest containing multiple 'assertions' — individual claims about the image's provenance.

A typical C2PA manifest from a DALL-E 3 image contains the following assertions: a 'c2pa.created' action assertion recording the creation timestamp and tool identifier, a 'c2pa.hash.data' assertion containing a cryptographic hash of the original pixel data, a 'stds.schema-org.CreativeWork' assertion with metadata about the work, and potentially an 'openai.dalle.generation' assertion with model-specific data. The entire manifest is signed with a digital certificate issued to the creating organization (OpenAI, Adobe, Microsoft, etc.).

Why Standard EXIF Tools Cannot Remove C2PA

The most common mistake creators make when trying to remove C2PA data is using a standard EXIF stripping tool. Tools like ExifTool, GIMP's 'Export without metadata' option, or online EXIF removers are designed to remove EXIF, IPTC, and XMP data — but they typically leave JUMBF containers intact. This is because C2PA was deliberately designed to be stored in a separate container that is not affected by standard metadata removal operations.

ExifTool, for example, can read C2PA data with the -C2PA flag, but its standard '-all=' metadata removal command does not delete JUMBF containers by default. You need to explicitly target the JUMBF data with a command like 'exiftool -JUMBF= -APP11= filename.jpg' — and even then, the behavior varies by ExifTool version and file type.

The most reliable method for complete C2PA removal is not to strip the JUMBF container from the existing file, but to re-encode the image from its pixel data — a process that creates a new file that never contained C2PA data in the first place.

The Canvas Re-encoding Method: Why It Works

The most technically sound approach to C2PA removal is to draw the image onto an HTML5 Canvas element and then export the canvas as a new image file. This works because the Canvas API operates exclusively on pixel data — it has no concept of file-level metadata containers. When you call canvas.toDataURL('image/jpeg', 0.92), the browser generates a brand new JPEG file containing only the pixel data from the canvas. No EXIF, no XMP, no IPTC, and critically, no JUMBF container with C2PA credentials.

This approach is not just effective — it is also the safest from a privacy perspective. Because the entire operation happens in the browser using the Canvas API, the original image never needs to leave your device. There is no server that receives your files, no database that logs your uploads, and no third party that could potentially access your original images.

Step-by-Step: How BlankAI Removes C2PA

1. 1You select your image(s) in the upload zone. The files are read into memory using the FileReader API — no network request is made.
2. 2Each image is decoded and drawn onto an off-screen HTML5 Canvas element at its original resolution.
3. 3A pixel modification algorithm iterates through the canvas pixel data, applying imperceptible ±1 RGB adjustments to selected pixels. This changes the SHA-256 hash of the pixel data, which invalidates any C2PA hash assertions that referenced the original pixel data.
4. 4The modified canvas is exported as a new JPEG at 92% quality using canvas.toDataURL(). This creates a completely new file with no metadata containers of any kind.
5. 5A new SHA-256 hash is computed for the output file using the Web Crypto API, allowing you to verify that the hash has changed.
6. 6The cleaned image is made available for download as a data URI — a format that iOS Safari can save directly to the Photos app.

Verifying That C2PA Has Been Removed

After removing C2PA credentials, you should verify that the removal was successful before using the image. There are several ways to do this. The Content Authenticity Initiative's own verification tool at verify.contentauthenticity.org will show 'No Content Credentials found' for successfully cleaned images. ExifTool with the -C2PA flag should return no results. And BlankAI's own Image Diff tool can compare the original and cleaned images, showing the hash change and confirming that metadata fields have been removed.

The Cryptographic Signature Caveat

One important technical nuance: C2PA's cryptographic signature means that the absence of credentials can itself be a signal in some contexts. If a platform knows that a specific AI tool always embeds C2PA credentials, and it receives an image that appears to have been generated by that tool but lacks credentials, it may flag the image as having had credentials removed. This is a relatively sophisticated detection method that requires the platform to have prior knowledge of the generating tool's behavior.

In practice, this detection method is not widely deployed in 2026, and it only applies to images from tools that always embed C2PA. For images from tools that do not embed C2PA (most Stable Diffusion deployments, many fine-tuned models), this concern does not apply. The pixel-level fingerprint modification that BlankAI applies also helps in this scenario, as it changes the image's statistical characteristics in ways that make it harder to definitively attribute to a specific AI tool.

Comparison: C2PA Removal Methods