What Is AI Metadata?
How generative models like Midjourney, Stable Diffusion, and DALL-E encode prompts, model versions, and generation seeds inside images.
When generative AI platforms produce images, they often write machine-readable parameter metadata directly into the exported file structure. This data is not visible when viewing the image visually, but can be extracted instantly with any metadata parser.
Common Data Fields Stored in AI Images
- Positive Prompt: The descriptive text entered by the user to produce the visual composition.
- Negative Prompt: Concepts, colors, or artifacts instructed to be excluded from the generation.
- Seed Number: The integer used to seed the initial random latent noise field.
- Model & Checkpoint: The specific architecture and weights used (e.g., SDXL 1.0, Midjourney v6).
- Sampler Parameters: The diffusion algorithm, steps, and CFG (Classifier-Free Guidance) scale.
Where Is This Metadata Stored?
The exact container depends on the image file format:
- PNG: Stored within auxiliary chunks such as
tEXtoriTXtwith keys likeparametersorprompt. - JPEG / WebP: Frequently stored inside the EXIF block under the
UserCommenttag or inside an embedded XMP packet.
Why Creators Remove AI Metadata
Many digital artists, designers, and agencies prefer to strip AI metadata to safeguard proprietary prompting methodologies, remove software lineage tags before client delivery, or prevent automated scraping.
Inspect or Clean Your AI Images
Process your files locally in your browser with zero server uploads.
Which AI tools write which metadata?
What a file carries depends on the tool and on how you exported or downloaded it. Use this as a starting point, then check your own file.
| Tool | What is usually stored | Where |
|---|---|---|
| Automatic1111 / Forge (Stable Diffusion) | Prompt, negative prompt, steps, sampler, CFG, seed, model | PNG text chunk named “parameters” |
| ComfyUI | Full workflow and prompt graph as JSON | PNG text chunks “workflow” and “prompt” |
| ChatGPT image generation and DALL-E | Signed Content Credentials stating AI generation | C2PA manifest |
| Adobe Firefly, Photoshop generative tools | Content Credentials and AI fields | C2PA manifest and XMP |
| Google Gemini and Imagen | An invisible watermark in the pixels | SynthID (not metadata) |
Why does AI metadata matter?
- Privacy: prompts can contain names, places or client details, and workflows can reveal how you work.
- Platform labels: Meta has said it reads C2PA and IPTC metadata to apply its “AI info” label.
- Reproducibility: the seed and settings let you recreate an image, which is why many artists keep a private original.
Frequently asked questions
Is AI metadata visible in the picture?
No. It is stored in containers around the image data (PNG text chunks, EXIF, XMP, C2PA), so the picture looks the same with or without it. A metadata viewer shows it.
Does every AI image contain AI metadata?
No. Many services strip it on export, and screenshots, messaging apps and some social platforms create new files without it. Check the specific file you plan to share.
Can AI metadata be removed?
Yes. Removing the containers that hold it leaves the pixels unchanged. It does not remove pixel watermarks such as SynthID.
Sources
- Digital Source Type vocabulary (IPTC)
- C2PA and Content Credentials Explainer (C2PA)
- Portable Network Graphics (PNG) Specification, Third Edition (W3C)
- Meta changes its label from “Made with AI” to “AI info” (TechCrunch, July 1, 2024)