DALL-E vs. MidJourney vs. Stable Diffusion: What Metadata Each Leaves Behind
Direct Technical Summary
Generative AI models store provenance in completely different container structures. Stable Diffusion (A1111 / ComfyUI) embeds uncompressed or deflate-compressed text chunks (tEXt, zTXt, iTXt) containing full prompts, seed numbers, and node graphs. DALL-E 3 embeds cryptographically signed C2PA Content Credentials inside JPEG APP11 marker segments. MidJourney may include generation parameters in EXIF/XMP fields or PNG text depending on export method. None of these formats are universal: web transfers or secondary edits often alter or strip these chunks unpredictably.
Comparative Breakdown: AI Generator Metadata Footprints
The table below illustrates the typical metadata structures observed in raw direct exports from the leading image synthesis engines:
| Generator | Primary Format | Container Marker / Chunk | Typical Content Embedded | Client-Side Removal |
|---|---|---|---|---|
| Stable Diffusion (A1111) | PNG | tEXt / zTXt ("parameters") | Prompt, negative prompt, seed, sampler, steps, CFG scale. | Lossless (Mode A) |
| ComfyUI | PNG | tEXt ("prompt", "workflow") | Complete JSON node graph with custom model checkpoints. | Lossless (Mode A) |
| DALL-E 3 (OpenAI / ChatGPT) | JPEG / WebP | APP11 (C2PA) / c2pa chunk | C2PA manifest, signing certificates, assertion manifests. | Lossless (Mode A) |
| MidJourney (Web Archive) | PNG / WebP | tEXt or EXIF UserComment | Job ID, prompt string, aspect ratio flags (--ar, --v). | Lossless (Mode A) |
Stable Diffusion: The Risk of Full Workflow Leaks
Among all generative engines, local Stable Diffusion installations pose the highest risk of accidental metadata leakage. By default, WebUIs like Automatic1111 and ComfyUI embed extensive diagnostic metadata to allow creators to recreate images:
-
Proprietary Prompt Exposure: The exact prompt text, artist weightings, and negative prompts are stored in plaintext inside the
parameterschunk. -
ComfyUI Workflow Graphs: ComfyUI embeds the complete serialized node graph in the
workflowchunk. Anyone who downloads the PNG can drag it into ComfyUI to reveal your exact custom nodes, LoRA paths, and private folder structures.
Using our dedicated PNG Metadata Remover, you can strip ancillary text chunks while keeping image dimensions, alpha transparency, and color fidelity bit-for-bit identical.
DALL-E 3 & C2PA: Signed Provenance Records
Unlike open-source models that store generation parameters in simple text chunks, commercial platforms like OpenAI implement the C2PA specification. When an image is created in ChatGPT or via the DALL-E 3 API, the system attaches a signed provenance container:
In JPEG files, this data resides inside an APP11 marker segment with the signature JP2 or JUMBF (JSON Universal Media Byte Format). This cryptographic structure binds the creation event to OpenAI's signing certificate.
Mode A in our AI Metadata Remover scans the JPEG marker stream and excises the APP11 segment cleanly, copying the compressed entropy scan (the image payload) directly without recompressing pixels.
Important Limitations & Caveats
When working with AI-generated media, it is important to remember:
- Variability by Version: Different generator versions or community forks may alter chunk keys or disable metadata writing entirely. Do not assume all AI images contain metadata.
- Platform Re-encoding: Sending an image via messaging apps or social networks often re-compresses the file and strips textual chunks automatically.
- Pixel Watermarks: Container cleaning does not alter invisible frequency-domain watermarks. For details, read our Technical Disclaimer.
Frequently Asked Questions
Do all AI-generated images contain embedded prompts and metadata?
No. Whether an image contains generation metadata depends entirely on the generator software, version, and export pipeline. For example, Stable Diffusion local interfaces (Automatic1111, ComfyUI) often embed full text prompts and node graphs by default, whereas web interfaces or social platform CDNs often re-encode images and strip textual metadata during transmission.
How does Stable Diffusion store generation prompts in PNG files?
Stable Diffusion web UIs typically store parameters in PNG ancillary text chunks: tEXt, zTXt (compressed text), or iTXt (international text). Common chunk keywords include 'parameters', 'prompt', and 'workflow', which can contain full JSON structures describing positive prompts, negative prompts, seed values, CFG scale, and sampler settings.
What type of metadata does DALL-E 3 embed?
DALL-E 3 outputs delivered through official OpenAI APIs and ChatGPT platforms typically incorporate C2PA Content Credentials. In JPEG files, this is stored in APP11 marker segments containing signed digital provenance manifests declaring the software origin.
Strip Prompts and AI Manifests Privately
Clean Stable Diffusion ComfyUI workflows, PNG tEXt chunks, and C2PA markers 100% in browser memory.