How to Rip Prompt Metadata Out of an Image Before You Publish

The phrase "rip prompt metadata" is the one AI artists actually use in forums. It means: pull the prompt and the model info out of a file before anyone else can. The risk is real — the prompt is the entire recipe for your image. If a competitor downloads the file, runs exiftool, and reads parameters, they get your exact seed, CFG scale, sampler, and prompt text. For a POD seller on Etsy or Redbubble, that is enough for someone to reproduce the asset in a minute and undercut your price.

Which generators leak the prompt?

Almost all of them, by default:

• Stable Diffusion (AUTOMATIC1111, ComfyUI, Forge, SD.Next) — full prompt, negative prompt, sampler, CFG scale, steps, seed, model hash, and LoRA hashes. All in the PNG tEXt / iTXt chunks, or in the XMP parameters field for JPEG exports. The full recipe.
• Midjourney — prompt text, model version (--v 6.1), aspect ratio, style tuner flags, and seed. In the XMP parameters block. Midjourney also stores a per-image UUID in the file name; the prompt itself is in the metadata.
• Adobe Firefly — no prompt in the file. The generator is identified by a C2PA Content Credential instead. This is the only major generator that does not leak the prompt.
• OpenAI DALL·E 3 — the prompt is in the dalle XMP field, but rewritten (DALL·E rewrites user prompts before generation, and the rewritten version is what is stored). Still extractable.
• Ideogram — prompt in the ideogram XMP field, model version included.
• Microsoft Copilot Designer / Bing Image Creator — C2PA credential only. No prompt leak.

If you are using anything other than Firefly or Copilot, your prompt is in the file.

How do you rip prompt metadata before publishing?

A byte-level cleaner. The workflow:

1. Generate the image in your usual tool.
2. Export as JPEG, PNG, or WebP (HEIC is read-only in most tools).
3. Drop the file into our free scanner to see exactly what the file carries — prompt text, model, seed, C2PA credential, all of it.
4. Click Clean. The cleaner walks the file, removes the XMP parameters block, the dalle field, the ideogram field, the PNG tEXt chunks with the prompt, and the C2PA JUMBF box. The pixel data is byte-identical to the source.
5. Re-scan to confirm the prompt is gone. The report should show zero AI-trailing fields.
6. Upload the cleaned file to Etsy, Redbubble, or wherever you list.

The whole flow runs in your browser. The file never leaves your device.

Why this matters for POD sellers specifically

Marketplaces require AI disclosure. Buyers can download your listing image, run a metadata scan, and see both the prompt and the model. If your listing copy says "Original hand-drawn art" and the file's parameters block says "studio ghibli, oil painting, dramatic lighting", the dispute is over before it starts.

A clean file with no prompt trail is not a defense against the disclosure rule — you still have to disclose that the asset is AI-generated. But it is a defense against a buyer extracting your recipe and undercutting you with their own version. Stripping the prompt does not change the disclosure obligation, but it does close the IP leakage door.

Can you keep the prompt and strip everything else?

Technically, yes. Most byte-level cleaners let you choose: strip all AI metadata, keep only the C2PA credential, keep only the XMP parameters, or keep the lot. If you want the audit trail but not the prompt, that is a one-click option. If you want the prompt kept for your own records, that is too — just download the original separately and store it outside the published file.

The default in our tool is to strip everything AI-related. The cleaning strategy guide walks through the per-format field-by-field decision.

One thing the prompt metadata does not tell you

Knowing someone's prompt does not let you recreate their exact image. The seed gives you the random-number state, the model gives you the weights, the CFG scale and steps constrain the diffusion process — but the renderer also depends on the runtime (PyTorch version, GPU, attention backend), the LoRA stacking order, and the exact preprocessing pipeline. Two artists using the "same" Stable Diffusion setup will produce different images from the same prompt and seed.

What the prompt metadata does tell you is the creative direction. A prompt like "isometric cozy coffee shop, soft lighting, pastel colors" is a recipe anyone can iterate on. Stripping the prompt does not protect the concept, but it does remove the convenience of one-click copy.

Related reading

• How to check if an image was AI-generated — the read-only side of the same workflow.
• POD sellers: AI metadata on Etsy, Redbubble, and Printify — disclosure rules and the marketplace scan workflow.
• Remove AI metadata from images — the hub page with the full byte-level workflow.
• What is a C2PA Content Credential? — the signed manifest that Firefly and Copilot write instead of a prompt.