Most discussion of image metadata is about EXIF — the camera, the lens, the GPS coordinates. Generated images carry something different and frequently more sensitive: the exact instructions used to make them. People share generated PNGs constantly without realising the file describes, in plain text, how it was produced and often the entire pipeline behind it.
What is in there
Generation tools write to PNG text chunks, not standard EXIF — so an EXIF-only check will report the file as clean when it is not.
Typical contents: full prompt, negative prompt, seed, sampler, steps, CFG, model hash, LoRAs. Node-based tools embed the entire workflow graph as JSON.
A real example
Read that as a competitor would. It reveals the phrasing that produces your house look, the fact that you maintain a custom style LoRA, your internal model naming, and the exact settings to reproduce the result. For a studio whose output style is the product, this is the recipe.
Node-based tools go further. A workflow graph embedded as JSON can run to tens of kilobytes and describes every node, connection, model and parameter in your pipeline — effectively your production process, shipped inside every image you publish.
⚠️ An EXIF check is not enough
Generation data lives in PNG tEXt and iTXt chunks, which are a PNG mechanism, not EXIF. Tools and online viewers that only read EXIF will report "no metadata found" on a file containing your entire prompt.
Check for both. If a viewer shows nothing on a file you know came straight from a generator, the viewer is the problem — the data does not usually vanish on its own.
When to strip, and when not to
This is a genuine trade rather than a default. The same metadata that leaks your process is what establishes where an image came from.
| Situation | Strip? |
|---|---|
| Prompt is commercially sensitive | Yes |
| Workflow reveals your pipeline | Yes |
| Client work under confidentiality | Yes |
| Publishing at volume on your own site | Yes — it is also dead weight |
| You want provenance to travel | No — keep C2PA |
| Disclosure is required or expected | No |
| Sharing so others can reproduce it | No — that is the point |
| Your own archive master | No — never |
The distinction that resolves most cases: generation parameters and provenance are different things. Your prompt is a private working detail. A C2PA manifest is a public claim about origin, and removing it makes an image less trustworthy rather than more private — it does not conceal that the image is generated, it just removes the evidence that anyone can check.
If you want both, strip the parameters and keep the manifest. That is a deliberate configuration rather than something any tool does by default.
🚨 Never strip your archive copy
The seed, prompt and settings are how you regenerate, extend or produce variations of an image later. Strip the master and that capability is gone permanently — you cannot recover a seed from pixels.
Strip derivatives, never originals. Keep the full-metadata PNG privately and publish cleaned copies, which is the same master-and-derivative pattern that handles format conversion.
How to remove it
Four approaches, in descending order of reliability:
- An explicit metadata stripper. Removes ancillary chunks and EXIF deliberately, leaves the pixels untouched. The right tool for the job — our EXIF remover does this in the browser, so the image never leaves your device.
- Format conversion. PNG to WebP or JPEG usually drops text chunks because there is nowhere to put them. Effective in practice, but it is a side effect rather than a guarantee, and it re-encodes your pixels.
- Re-encoding through a processing step. Resizing or compressing typically produces a fresh file without the original chunks. Same caveat — incidental, and tool-dependent.
- Screenshot. Crude, lossy, guaranteed. Nothing survives because you are capturing rendered pixels. Worth knowing as a last resort.
Verify — do not assume
The step almost everyone skips. Tools vary, options change, and a pipeline that stripped metadata last month may not after an update.
Three checks worth building into a routine:
- Search the bytes for an unusual term from your prompt or your model name. Text chunks are stored uncompressed in many cases, so this finds them directly.
- Open the cleaned file in a viewer that reads PNG chunks, not just EXIF.
- Watch the file size. An embedded workflow graph can be tens of kilobytes; if a "stripped" file is the same size as the original to the byte, be suspicious.
💡 Platforms strip metadata — but on their schedule, not yours
Most social platforms remove metadata on upload, so a prompt shared there is usually gone. That is not a control you own: it varies by platform, by upload path, and it does not apply to files sent by email, shared in cloud storage, or posted anywhere that preserves originals.
Direct file sharing is where leaks actually happen, because that is the path where the original survives intact. Strip before the file leaves your machine and the question never depends on someone else's behaviour.
Summary
- Generators write to PNG text chunks, not EXIF — EXIF-only checks miss everything.
- Contents include prompt, negative prompt, seed, model and LoRAs, and sometimes a full workflow graph.
- For a studio, this is the recipe, not trivia.
- Parameters and provenance are different. Strip the first, consider keeping the second.
- Never strip your archive master — the seed cannot be recovered from pixels.
- Strip deliberately, then convert. Do not rely on conversion as the removal step.
- Verify by searching the raw bytes for a distinctive prompt word.
- Platform stripping is not your control. Direct file sharing is where leaks happen.
Frequently Asked Questions
What metadata do AI image generators embed?
Commonly the full prompt and negative prompt, the seed, sampler, step count, guidance scale, model name or hash, and any LoRAs used. Node-based tools often embed the entire workflow graph as JSON, which describes your whole pipeline. This lives in PNG text chunks rather than standard EXIF.
How do I remove metadata from an AI-generated image?
Re-save the image without its ancillary chunks, using a tool that strips metadata explicitly rather than relying on a side effect. Converting to another format usually drops it too, but that is incidental behaviour that varies by tool, so verify the result rather than assuming.
Does converting PNG to JPEG remove the prompt?
In most cases yes, because PNG text chunks have no equivalent in JPEG and get discarded. It is not a guarantee — some tools copy metadata across formats where they can — so treat conversion as a likely side effect rather than a deliberate removal, and check the output.
Should I always strip metadata from AI images?
No. It is worth removing when the prompt is commercially sensitive, when workflow details reveal a pipeline you would rather keep private, or for client work under confidentiality. But metadata is also how provenance works, so stripping a C2PA manifest removes a signal that says where the image came from.
How can I check whether metadata was actually removed?
Inspect the file after processing rather than trusting the tool. Open it in a metadata viewer, or search the raw bytes for a distinctive word from your prompt — if the text appears, the metadata is still there. Comparing file sizes before and after is a weak signal but a useful first hint.