Remove AI Metadata From Your Images

What AI generators actually write into your files

Each toolchain leaves its own fingerprint. ChatGPT images ship with a C2PA Content Credentials manifest — a signed record naming the generator and creation details, stored in a dedicated section of the PNG or WebP container. Gemini images carry metadata labels alongside Google's invisible SynthID watermark. Adobe Firefly attaches Content Credentials through the same C2PA standard. And Stable Diffusion web interfaces are the most talkative of all: many write your full prompt, negative prompt, seed, sampler and model name into a plain-text chunk of the PNG.

None of this shows up when you look at the image. All of it ships when you share the file.

Why people strip AI metadata

The prompt is the clearest case: if you've iterated on a prompt for client work or a product you sell, a Stable Diffusion PNG can hand your entire recipe to anyone who inspects the file. Provenance manifests are subtler — they record which tools you used, when, and can carry identity signals, which matters if you publish under a pseudonym or deliver assets a client will redistribute. And AI-detector websites read exactly this metadata layer, so a file destined for a marketplace or portfolio may be flagged purely on fields you didn't know were there.

There are also mundane reasons: some platforms mangle files with unfamiliar metadata blocks, and cleaned files are simply smaller.

What this tool removes from AI images

The full metadata layer, whatever wrote it: C2PA Content Credentials manifests (the JUMBF blocks in JPEG APP11 segments, PNG caBX chunks and WebP C2PA chunks), XMP packets with AI-generation labels, EXIF blocks, IPTC records, and every PNG text chunk — which is where Stable Diffusion prompts live. The per-file report itemises each removed block and its size before you download.

The strip is structural and lossless: metadata sections are deleted whole and your pixel data is copied through byte for byte, never decoded or re-encoded. The cleaned image is visually identical and a little smaller.

What removal does not do — read this part

Metadata is only one of the ways AI images are labelled. Google's SynthID and similar systems embed an invisible watermark into the pixel pattern itself, designed to survive editing, cropping and metadata removal — this tool does not touch it, and detectors built on pixel-level watermarks will still identify those images as AI-generated. Removing metadata makes a file quieter, not anonymous.

Our position is the same one we take on the Gemini watermark page: cleaning your own files for your own use is a legitimate privacy and workflow choice. Passing AI output off as human work, photographic evidence, or someone else's art is a different act, and metadata removal won't make it work anyway.

Check before, verify after

Drop an AI image in and read the report first — seeing your own prompt sitting in a text chunk is usually the moment this becomes concrete. Clean it, then drop the downloaded copy back in: the report should come back empty. Every step runs locally in your browser tab; open DevTools and watch the Network panel if you want proof that nothing is uploaded.