What's Really Inside Your AI Image -- DALL-E vs Midjourney vs Stable Diffusion Metadata

When you generate an image with an AI tool, the pixels are only part of the output. Beneath the visible image lies a layer of metadata that varies dramatically depending on which platform you used. DALL-E 3 embeds dozens of cryptographically signed provenance fields. Midjourney records software identifiers in EXIF and IPTC blocks. Stable Diffusion writes your entire prompt and generation parameters into the file. Understanding what each platform embeds -- and how much of it identifies the image as AI-generated -- is essential for anyone who creates, edits, or publishes AI imagery.

DALL-E 3: The Most Metadata-Heavy Platform

DALL-E 3, accessed through OpenAI's API and ChatGPT interface, produces images with the richest metadata profile of any current AI image generator. This is primarily because OpenAI has fully adopted the C2PA standard, making DALL-E 3 one of the first consumer-facing AI tools to embed cryptographically signed content credentials.

When you download a DALL-E 3 image, here is what is embedded:

C2PA Manifest Store: The image contains a complete C2PA manifest with over 47 assertion fields. The claim generator is identified as "OpenAI Media Service API" with a specific version number. The action type is "com.created" with a sub-type of "com.ai.generated" -- an unambiguous declaration that the content is AI-generated.

XMP Block: Beyond C2PA, DALL-E 3 embeds an XMP metadata block that includes additional OpenAI-specific fields. This typically contains the image generation API endpoint, a reference to the OpenAI terms of service, and a content identifier.

EXIF Data: Standard EXIF fields are populated, including the software field (set to "OpenAI DALL-E 3"), the creation date, and image dimensions. Unlike camera photos, no camera-specific fields are populated -- which itself is a signal that the image was not captured by a physical camera.

Prompt Information: In some configurations, the text prompt used to generate the image is embedded within the C2PA ingredient assertions. This means your creative input -- potentially containing proprietary concepts, client information, or sensitive descriptions -- is permanently stored inside the file.

The total metadata overhead for a DALL-E 3 image typically ranges from 20 KB to 35 KB, with the C2PA manifest store accounting for the majority of that size. The cryptographic certificate chain alone can contribute 8-12 KB.

Midjourney v6: EXIF and IPTC Tagging

Midjourney takes a different approach to metadata. As of v6, Midjourney does not implement C2PA content credentials. Instead, it relies on traditional EXIF and IPTC metadata standards to record information about generated images.

When you download an image from Midjourney (via the website or Discord), here is what you will find:

EXIF Data: The Software field is set to "Midjourney" along with a version identifier. The DateTime field records when the image was generated. Image dimensions and resolution are recorded in standard EXIF fields. Notably, Midjourney also writes a unique image identifier into the ImageDescription or UserComment EXIF field, which can be used to trace the image back to a specific generation job.

IPTC Data: Midjourney populates several IPTC fields, including the Caption field (which may contain a truncated version of the prompt), the Creator field (set to the user's Midjourney handle), and the Software field. The IPTC data also includes copyright information referencing Midjourney's terms of service.

No C2PA: The absence of C2PA content credentials means that Midjourney images are not automatically flagged by C2PA-aware platforms. However, the EXIF Software field and IPTC Creator field are easily read by basic metadata scanners, and many social media platforms check these fields to identify AI-generated content.

Midjourney's metadata footprint is significantly smaller than DALL-E 3's. A typical Midjourney v6 image contains 4-8 KB of metadata, primarily in EXIF and IPTC blocks. This makes Midjourney images lighter but also less cryptographically protected -- the metadata can be modified or removed more easily than C2PA-signed data.

Stable Diffusion: Prompt and Parameter Transparency

Stable Diffusion, particularly when run through popular interfaces like AUTOMATIC1111 WebUI or ComfyUI, takes the most transparent approach to metadata. The entire generation process -- prompt, negative prompt, sampler, steps, CFG scale, model hash, and more -- is written directly into the image file.

Here is what a typical Stable Diffusion WebUI image contains:

EXIF UserComment Field: This is where the magic happens. The WebUI writes a comprehensive JSON-formatted string into the EXIF UserComment field that includes: the full positive prompt, the full negative prompt, the sampler name (e.g., "DPM++ 2M Karras"), the number of steps, the CFG scale, the seed, the model hash, the VAE hash, the clip skip value, and the generation timestamp.

PNG tEXt Chunk: For PNG outputs, the same information is stored in a tEXt chunk with the key "parameters". This is even more accessible than EXIF data since PNG tEXt chunks are trivially readable by any image processing library.

No Cryptographic Protection: Unlike DALL-E 3's C2PA signatures, Stable Diffusion metadata is unsigned and unencrypted. It is plain text embedded directly in the file. Anyone can read it, modify it, or remove it with basic tools.

The metadata size for a Stable Diffusion image depends heavily on the prompt length. A typical image with a moderate-length prompt contains 3-6 KB of metadata. Images with very long prompts (particularly those using detailed negative prompts) can contain 10 KB or more of text metadata.

Real-World Impact: How This Metadata Affects You

The metadata differences between platforms have concrete consequences for creators and publishers:

Platform detection: Instagram and Facebook read C2PA content credentials automatically. A DALL-E 3 image with C2PA assertions will be labeled "AI-generated" immediately upon upload. A Stable Diffusion image without C2PA data may slip through undetected -- unless the platform also scans EXIF fields for known AI tool signatures, which Meta has begun doing.

Prompt leakage: A designer using Stable Diffusion for client work shared images directly from the WebUI without stripping metadata. The client extracted the full prompt from the EXIF data -- including negative prompts and parameter tuning that revealed the designer's proprietary workflow. This is a common and preventable mistake.

Chain of custody in legal contexts: In a 2025 copyright dispute, C2PA provenance data from a DALL-E 3 image was admitted as evidence to establish the image's creation date and generator. The cryptographic signing made the provenance data difficult to dispute. Without C2PA, establishing AI generation for legal purposes relies on less authoritative metadata like EXIF Software fields, which can be easily fabricated.

Content licensing: Stock photography platforms like Shutterstock and Adobe Stock require AI-generated content to be clearly labeled. Images with C2PA assertions are automatically identified; images without them may require manual disclosure. Failure to disclose AI generation on these platforms can result in account suspension and revenue clawback.

How to Check What Metadata Your Images Contain

Before sharing or publishing an AI-generated image, you should verify what metadata it carries:

1. ExifTool (command line): The gold standard for metadata analysis. Run exiftool -a -G1 image.png to see every metadata field, including C2PA assertions.

2. C2PA Validator (online): The C2PA project provides a web-based tool at contentauthenticity.org that displays the full provenance chain of any C2PA-enabled image.

3. PhotoME (Windows): A free GUI tool that displays EXIF, IPTC, and XMP data with a user-friendly interface.

4. Browser developer tools: Right-click an image on a website, inspect the element, and check the file's properties. This will not show C2PA data but will reveal basic EXIF and XMP fields.

5. RemoveAI Image (browser): After processing an image through RemoveAI Image, you can compare the before and after file sizes to confirm metadata removal. The difference is your metadata footprint.

FAQ

Which AI image generator embeds the most metadata?

DALL-E 3 embeds the most metadata by a significant margin. With over 47 C2PA assertion fields, XMP blocks, and EXIF data, a typical DALL-E 3 image carries 20-35 KB of metadata. Midjourney v6 is next with 4-8 KB, followed by Stable Diffusion with 3-10 KB (highly dependent on prompt length). The C2PA adoption by OpenAI and Google makes their images the most heavily tagged.

Can I selectively remove some metadata but keep others?

Yes. Metadata removal tools can target specific blocks. You could, for example, strip C2PA and XMP data while keeping basic EXIF fields like creation date and image dimensions. However, partial removal can itself be a signal -- a forensic analyst might notice that C2PA data is missing from a file format that typically includes it, which could raise suspicion. Complete removal that leaves the image with a clean, camera-like metadata profile is generally more effective.

Does uploading to social media remove AI metadata?

It depends on the platform and the type of metadata. Most social media platforms strip some EXIF data during upload (particularly GPS coordinates), but many preserve or even read C2PA data before stripping it. Instagram, for example, reads C2PA assertions to apply "AI-generated" labels and then may strip the C2PA data during re-encoding -- but the label remains. The key point is: platforms read your metadata before they strip it, and they act on what they find.

---

The metadata inside your AI images tells a story you may not want shared. Whether it is DALL-E 3's comprehensive C2PA provenance chain, Midjourney's EXIF and IPTC identifiers, or Stable Diffusion's complete prompt and parameter dump, this data follows your images everywhere. RemoveAI Image strips all of it -- C2PA, EXIF, XMP, IPTC, and GPS -- directly in your browser, with zero uploads to any server. Take control of what your images reveal.