DALL-E vs Midjourney vs Stable Diffusion: A Metadata Forensics Comparison

The Metadata Arms Race

Each major AI image generator handles metadata differently — from Midjourney's relatively minimal footprint to DALL-E 3's comprehensive provenance chain. Understanding these differences is critical for anyone working with AI-generated images, whether you're a creator, a platform engineer, or a digital forensics analyst.

This analysis examines the exact metadata structures embedded by each tool as of 2025.

DALL-E 3 (OpenAI)

What DALL-E Embeds

DALL-E 3, whether accessed through ChatGPT or the API, produces some of the most comprehensively tagged AI images:

EXIF Fields:

• Software: "DALL-E 3" or "OpenAI"
• ImageDescription: Contains the generation prompt (often truncated)
• Artist: "OpenAI"
• UserComment: May contain additional generation context

C2PA Manifest: DALL-E 3 is among the most aggressive adopters of the C2PA standard. Every image includes a cryptographically signed manifest that records:

• The generation tool (DALL-E 3)
• The generation timestamp
• OpenAI as the signing authority
• A tamper-evident content hash

XMP Data:

• xmp:CreatorTool: "DALL-E 3"
• dc:creator: "OpenAI"
• IPTC DigitalSourceType: "trainedAlgorithmicMedia"

Detection Difficulty: Easy

DALL-E 3's metadata is extensive and multi-layered. Even if EXIF data is stripped, the C2PA manifest provides a secondary detection vector. The C2PA data is embedded as a JUMBF (JPEG Universal Metadata Box Format) box that many simple metadata strippers miss.

Removal Strategy

Complete removal requires stripping both standard EXIF/XMP metadata and the C2PA JUMBF box. AI Metadata Cleaner's "Strip All Metadata" function handles both layers.

Midjourney

What Midjourney Embeds

Midjourney takes a comparatively lighter approach to metadata:

EXIF Fields:

• Software: Not consistently set — depends on the download method
• ImageDescription: Usually absent or minimal
• Artist: Not typically set

Steganographic Watermark: Midjourney embeds an invisible watermark directly into the pixel data using a proprietary steganographic algorithm. This watermark:

• Survives metadata stripping
• Survives moderate compression (up to ~70% JPEG quality)
• Survives moderate resizing (up to ~50% dimension change)
• Is detectable by Midjourney's own verification tool and some third-party detectors

XMP Data:

• Limited XMP data compared to DALL-E 3
• May include generation parameters when downloaded from the Midjourney website

Detection Difficulty: Moderate

Midjourney's metadata footprint is smaller, but the steganographic watermark adds a persistent detection vector that cannot be removed through metadata cleaning alone. Significant image manipulation (heavy editing, cropping, re-encoding at different quality levels) can degrade the watermark.

Removal Strategy

Metadata cleaning removes the identifiable EXIF and XMP fields. The steganographic watermark requires pixel-level modification (such as applying filters, adjusting levels, or re-encoding through multiple quality levels) which is beyond the scope of metadata tools.

Stable Diffusion (and Derivatives)

What Stable Diffusion Embeds

Stable Diffusion's metadata behavior varies significantly based on the frontend (WebUI, ComfyUI, Fooocus, etc.):

Automatic1111 WebUI:

• Parameters PNG chunk: Contains the complete generation recipe — prompt, negative prompt, sampler, CFG scale, seed, model name, and all generation settings
• Software: Usually "Automatic1111" or "Stable Diffusion"
• No C2PA manifest (open-source ecosystem hasn't widely adopted it)

ComfyUI:

• prompt PNG chunk: Contains the entire node graph as JSON — every node, connection, and parameter
• workflow PNG chunk: Contains the visual workflow definition
• This is by far the most verbose metadata of any AI tool

Fooocus:

• Parameters chunk: Similar to A1111 but with Fooocus-specific fields
• Software: "Fooocus"
• May include style and preset information

SDXL via API:

• Minimal metadata — depends entirely on the integration
• Some API wrappers add their own metadata; others produce clean files

Detection Difficulty: Variable

The detection difficulty ranges from trivial (ComfyUI with full workflow data) to challenging (API outputs with no metadata). The key distinction is that Stable Diffusion is open-source, meaning:

1. Users have full control over what metadata is embedded
2. Many frontends offer built-in options to disable metadata embedding
3. No centralized steganographic watermarking exists

Removal Strategy

For PNG files from A1111/ComfyUI: Removing the tEXt chunks containing Parameters, prompt, and workflow eliminates the generation data. Re-saving as JPEG also strips PNG-specific chunks.

For JPEG exports: Standard EXIF/XMP cleaning is sufficient.

Comparative Summary

*OpenAI has indicated future plans for pixel-level watermarking, but as of 2025 it is not deployed in production.

Key Takeaways

1. DALL-E 3 leaves the most comprehensive metadata trail but does not use steganographic watermarks, making it fully cleanable through metadata tools
2. Midjourney has a lighter metadata footprint but uses invisible watermarks that survive metadata cleaning
3. Stable Diffusion outputs vary wildly based on the frontend, but being open-source gives users maximum control over metadata
4. C2PA is the emerging standard, but currently only DALL-E 3 and Adobe Firefly consistently implement it
5. No single detection method is foolproof — platforms use combinations of metadata, visual analysis, and hash databases