How to Remove an AI Watermark From an Image: Metadata vs Pixels, Honestly
Updated September 2026
Most AI image watermarks are not in the image. They are in a signed block of metadata riding alongside it, and that block comes off in seconds with no loss. A minority of generators also embed a signal in the pixels themselves, and that one does not come off. Knowing which you have is the whole job.
Quick answer
Three steps:
- 1Scan the image. StripShot lists C2PA, XMP, IPTC, EXIF and any prompt text, and names the generator if it can.
- 2Strip. The metadata layer is removed at the binary level. Pixels are byte for byte the same.
- 3Read the re-scan. If the generator was Google, the page tells you SynthID is still in the pixels. Nothing else pretends otherwise.
Which generators embed what
| Generator | Metadata layer | Pixel layer | After StripShot |
|---|---|---|---|
| DALL-E 3 / ChatGPT images | C2PA manifest, XMP DigitalSourceType | None known | Clean |
| Adobe Firefly, Photoshop generative features | C2PA manifest, XMP | None known | Clean |
| Midjourney | EXIF and XMP description fields | None known | Clean |
| Stable Diffusion, ComfyUI, Automatic1111 | PNG tEXt parameters and workflow chunks, EXIF UserComment | None | Clean, prompt gone |
| Flux, Leonardo, Ideogram | XMP and C2PA depending on the host app | None known | Clean |
| Google Imagen, Gemini image output | C2PA, IPTC DigitalSourceType | SynthID | Metadata clean, SynthID remains |
| Grok (xAI) | XMP and C2PA depending on export | None known | Clean |
"None known" means none published and none observed. Vendors can add a pixel watermark at any time, which is one more reason to scan the actual file rather than trust a table.
How the metadata layer is removed
- JPEG. The file is a sequence of segments. Metadata lives in APP1 (EXIF and XMP), APP13 (IPTC) and APP11 (C2PA JUMBF boxes). StripShot walks the segments and copies only the ones that describe pixels, plus the JFIF header.
- PNG. A sequence of chunks. tEXt, iTXt and zTXt carry prompts and tool names, eXIf carries EXIF, caBX carries C2PA. Only the chunks a decoder needs are kept.
- WebP. EXIF, XMP and ICCP chunks are dropped and the VP8X flags are cleared so viewers stop looking for them.
- HEIC and AVIF. Offsets are absolute, so metadata boxes are neutralized in place and payloads zeroed. Same size, same pixels.
In every case the compressed image data is not decoded, which is why the output is pixel-identical to the input and often smaller.
The pixel layer, in one paragraph
SynthID and similar marks spread a signal across the whole image so it survives cropping, compression and screenshots. There is no block to remove. Research methods that regenerate the image with a diffusion model can weaken the mark at the cost of changing the picture, and nobody outside the vendor can confirm the result. StripShot does not offer this and does not claim to. The SynthID explainer has the detail.
After you strip
Check the re-scan, then post. If the platform later re-encodes the image, which most do, nothing you removed comes back. If you edited the image in Photoshop with a generative feature on the way, scan again: Adobe writes a fresh C2PA manifest on save, even into a photo you took yourself.
Free image watermark remover
Drop an AI image to see and strip its metadata
Runs in your browser. Nothing is uploaded.
Need unlimited strips?
Scans are free. Pro is unlimited.
Images, video, audio, documents, PDF and text in one tool. Pro adds unlimited strips, whole-folder batches and zip download with a JSON report.
Frequently asked questions
How do I remove the AI watermark from an image?
If the watermark is metadata, which is the case for DALL-E, Midjourney, Firefly, Flux, Leonardo and Stable Diffusion, drop the image into StripShot and click Strip. The C2PA manifest, XMP tags and EXIF are removed at the binary level with no re-encoding. If the image is from Gemini or Imagen, the same step removes the metadata but SynthID in the pixels remains.
Does a screenshot remove an AI watermark?
It removes metadata, because a screenshot is a new file with none. It does not remove a pixel watermark, which is designed to survive screenshots, and it costs you resolution and adds the screenshot device's own metadata.
Which AI image generators use pixel watermarks?
Google's Imagen, Gemini image output and Veo carry SynthID. Most others rely on C2PA and XMP metadata alone as of this writing. Vendors change policies, so scan the file rather than assume.
Will removing the metadata stop the Made with AI label on Instagram?
For images whose only provenance is metadata, yes, because Meta reads C2PA to apply the label. For Google-generated images, the label trigger is removed but Meta could still detect SynthID if it chooses to check.
Does removing the watermark reduce image quality?
Not with StripShot. Metadata segments are removed from the container and the compressed pixel data is copied unchanged. Tools that re-encode through a canvas or convert to JPEG do lose quality.