Remove SynthID Watermark: What Works and What Doesn't
SynthID is Google DeepMind's invisible watermark — a statistical signal woven into the pixels of AI images from Gemini, Nano Banana, Nano Banana Pro, Imagen and ChatGPT. It is not metadata, and that is the catch most "SynthID removers" miss. Stripping EXIF or C2PA tags leaves the file looking clean while the SynthID signal stays baked into the image, ready to be read straight back by Google's detector.
If you are here because an upload got labelled, this is why. Instagram, TikTok and X read these markers automatically and tag the post as AI-generated, and a labelled post reaches fewer people than an unlabelled one. There is no platform setting that turns the label off while the file still carries the signal, so the fix has to happen before you upload — the platform guides cover how each one behaves.
You will also find pages arguing that SynthID cannot be removed at all. On the evidence they cite, they are right — about editing. Cropping, filtering, screenshotting and re-compressing an image do not defeat SynthID, because the signal is spread redundantly across millions of pixels. The fix has to happen in the pixels, because that is where the signal is. That is a different and harder operation than deleting a tag from the file header, and the section below explains why that distinction decides whether a tool can help you at all.
For AI video, in-pixel SynthID removal is not reliable yet — so video mode removes the C2PA content credentials only, and we say so up front instead of pretending otherwise.
Can SynthID Actually Be Removed?
Not by editing the image — the research on that is sound. It can be cleared by regenerating the image, which is a different operation entirely.
The pages telling you SynthID is unremovable are describing a real property of the watermark, and it is worth understanding rather than dismissing. SynthID is not stored in any one place you can find and delete. It is a low-amplitude statistical pattern distributed redundantly across the whole image, so no single pixel or region carries it. Destroy part of the image and enough of the pattern survives in the rest to still read.
That is why every surface-level approach fails. Blur a section, clone over an area, apply a filter, re-save at a lower quality — the signal is spread too wide and too deep for any of it to matter, and the edits aggressive enough to disrupt it are aggressive enough to visibly wreck the picture. If your mental model is "find the watermark and erase it," the skeptics are correct and there is no surgical option.
The signal is in the picture, so the picture is what has to be processed. That is the whole difference between this and a metadata cleaner. A C2PA manifest is a block of data in the file header, and deleting it is trivial. SynthID is a pattern distributed across the pixels themselves, so there is nothing to delete — the image has to be put through processing that leaves the signal unreadable while leaving the picture usable.
That trade-off is the entire difficulty, and it is why this runs on our servers with a credit attached rather than instantly in your browser. Anything blunt enough to destroy a signal spread redundantly across millions of pixels is blunt enough to be visible, and a cleaned file is only worth having if it is still the image you made. Judge the result the way you would judge any edit: open it next to the original and look at the details you care about.
The obvious question is why you could not just run an image model over it yourself. The gap is fidelity. A general-purpose generator handed your picture returns a lookalike — the same subject, a visibly different image, useless as a replacement for your file. Holding the output close enough to the original that the two are hard to tell apart is the part that is difficult, and it is what you are paying for.
One honest caveat we will not bury: defeating the vendor's own verifier is not the same as being invisible to every forensic classifier. Published research finds that watermark-removal pipelines can leave signatures other detectors are able to spot. What this clears reliably is the provenance check platforms actually run at upload time. We do not promise universal undetectability, and you should distrust any tool that does.
What SynthID Survives — and Why That Matters
Every ordinary edit people try first leaves the watermark readable. This is the table worth checking before you pay anyone.
SynthID was engineered specifically to survive the manipulations a person would reach for. Robustness varies with how aggressive the change is, but the pattern is consistent across published testing:
| What you do to the image | Watermark still readable? |
|---|---|
| Screenshot it | Yes — minor confidence loss only |
| Crop it moderately | Yes |
| Crop away most of the frame | Often no — but you have lost the image |
| Resize or upscale | Yes |
| Adjust brightness, contrast or colour | Yes |
| Convert PNG to JPEG | Yes |
| Re-compress at normal quality | Yes |
| Delete all metadata | Yes — metadata is a separate layer |
| Process the pixels (hidden-watermark mode) | No — this is what we do |
The practical reading: if a tool describes its method as stripping, cleaning, filtering or optimising your file, it is operating in the rows that say "yes." Ask what the method is before you pay, and verify the result with a detector rather than trusting a success message — SynthID is invisible, so you cannot confirm anything by looking at the image.
Metadata Stripping Is Not SynthID Removal
Deleting EXIF and C2PA removes the "made with AI" labels but leaves the pixel signal completely intact.
Most free "SynthID removers" only delete metadata — EXIF, IPTC, and sometimes the C2PA manifest. That clears a label when the label came from the file header, and it does nothing to SynthID itself. You can run a metadata cleaner, upload the result to Google's SynthID Detector, and still get "watermark detected."
This matters most in the case people actually hit: a platform labelled your post, you found a metadata tool, the label came back on the next upload. That is the signature of a pixel-based trigger being treated as a header problem. Every competitor page in this category carries some version of the same admission — that if the label survives their tool, the cause was SynthID and they cannot help.
Here, both layers are handled in one pass. The cleaned image no longer carries a readable SynthID pattern, and because the result is written fresh from pixel data, it carries no C2PA manifest from the original generator either. One operation, both layers, one charged unit.
Which AI Images Carry SynthID
Gemini, Nano Banana, Nano Banana Pro, Imagen, and — since OpenAI adopted it in 2026 — ChatGPT images too.
Google applies SynthID across its image stack: Gemini (including the Nano Banana and Nano Banana Pro image models), Imagen, and images generated through the Gemini app, AI Studio, and Vertex AI. OpenAI's ChatGPT images (GPT Image and DALL·E 3) now carry Google's SynthID in the pixels too — OpenAI adopted it in 2026 — alongside their C2PA Content Credentials. If your image came out of any of these, a plain metadata wipe will not clear the hidden watermark.
This page handles all of them. Not sure your image is even marked? You can check it for SynthID and C2PA first — worth doing, because if your file only carries C2PA you do not need the credit-consuming pass at all. Otherwise upload the file, choose hidden-watermark removal, and download the clean result.
SynthID on Video: Why We Only Remove C2PA
We do not remove in-pixel SynthID from video, and we would rather tell you than sell you a metadata wipe under another name.
Google extends SynthID to AI video from Veo and Flow, embedding the signal across frames. Regenerating every frame of a clip at the fidelity a still image gets is a materially harder problem — and any tool advertising fast, lossless "SynthID video removal" is almost certainly deleting metadata and calling it SynthID.
So here is the limit, stated plainly. In video mode this tool removes the C2PA content credentials and container provenance metadata — the layer that marks a clip as AI-generated for YouTube, TikTok, Instagram, Adobe and stock libraries — while copying the encoded video and audio untouched, with no quality loss. The in-pixel SynthID signal on video is not removed.
If your clip was labelled and the C2PA strip does not clear it, the trigger was pixel-based and we cannot fix that yet. For a full clean, use image mode on a still frame.
Quality, Privacy, and How It Compares
Visually identical output, files kept in your account only, and credits charged solely on success.
Regeneration is tuned to stay below the threshold of human perception — no blur, banding, or colour shift, and no visible softening of detail. Video mode copies the encoded streams with no re-encode, preserving source resolution, framerate, and audio exactly. Standard mode uses 1 credit and HD mode uses 2, charged only when a job succeeds; if a job fails the reservation is released and nothing is deducted.
Where thin competitors offer a one-line "drag and drop, metadata gone" widget, this tool does the part that actually matters and tells you precisely what each mode does and does not clear. Your files are stored in your own account, never used for model training, and downloaded through time-limited signed links. For the visible Gemini sparkle logo, use the main Gemini image tool; for ChatGPT stills, the ChatGPT image remover handles SynthID and C2PA the same way. Real results are on the before and after page.