Two different things share one name
When someone asks about removing an AI watermark, they are usually asking about a badge they can see in the corner of a generated image. That is the easy case, and it is solvable. But the same phrase also covers something much less tractable, and conflating the two is the source of most disappointment in this category.
1. The visible overlay — removable
A semi-transparent logo drawn on top of the image. Underneath it, the original pixels are still there. The mark was blended in, not carved out, which means the operation can be reversed exactly.
2. The invisible signal — not removable
An imperceptible fingerprint embedded in the media itself. Nothing was layered on top, so there is no layer to take off. Google calls this SynthID and states it is built to survive re-scaling, re-colouring and compression.
3. Signed provenance — not a watermark to remove
C2PA Content Credentials are signed metadata recording an edit history. They are not an overlay, and tampering with them invalidates the signature rather than removing anything.
Before you remove anything: Google’s Terms of Service state you must not remove, obscure or alter their branding, logos or legal notices. That obligation can apply to your own output. Check the sourced legal analysis before using any of this commercially.
Which models add what
| Model | Visible overlay | Invisible signal | Removable here |
|---|---|---|---|
| Gemini (chat & apps) | Yes | SynthID | Overlay only |
| Nano Banana (Gemini 2.5 Flash Image) | Yes | SynthID | Overlay only |
| Nano Banana Pro / 2 (Gemini 3.1 Flash Image) | Yes | SynthID | Overlay only |
| Veo, Veo 3, Veo 3.1 | Yes | SynthID | Overlay only |
| Sora | Configurable | Provider-specific | Not handled |
Google AI Studio states directly that all Nano Banana images include an invisible SynthID watermark, and that Google can check a file for you and report which parts of a video carry it. That is the right tool for the question “does this file have a mark”, and it beats asking a remover.
Because every Google model stamps an identical overlay, one tool covers the whole range. There is no need for a separate Nano Banana remover or a Veo remover; the watermark being removed is the same watermark. The per-model detail is broken out in the full comparison.
How to actually remove the visible overlay
- Use the original file. The single biggest cause of failure. A screenshot, a re-save, a crop, or an upload through a messaging app re-encodes the file and destroys the recoverable signal, while leaving the mark plainly visible.
- Open Image Studio or Video Studio. Both run entirely in your browser, so the file is never uploaded.
- Check the detected region matches what you see. A low-confidence result almost always means the wrong file.
- Export and inspect. On video, check the first and last frames and confirm audio still plays.
Why most removers blur, and why this one does not
Most tools treat the marked area as damaged and repair it with generative inpainting. That is solving the wrong problem. The model has no access to the pixels you want — they are one layer down — so it infers a plausible continuation of the surrounding texture. Plausible is not the same as correct, and it shows wherever the content has fine structure.
The alternative is arithmetic. A visible mark is composited with standard alpha blending, so the composite is a known function of the mark, the original and the opacity. Solve it backwards and the original is recovered exactly:
Nothing is synthesised, so the smearing failure mode does not exist. The harder engineering problem is measuring the geometry and opacity profile accurately, which is covered in how the overlay is detected, and the honest limits are in removing a watermark without blur.
Why the model name is a clue, not a proof
It is natural to assume the answer follows from which model made the file, and sometimes it does. But defaults change, several of these models carry a visible toggle, and the same family ships generations with different behaviour. A reliable check is faster than a wrong assumption, and it takes about a minute.
- Look for something visible. Point at it. If there is a badge you can see, that is the recoverable layer, and this tool addresses it. If there is nothing visible, an overlay remover has nothing to do, and a tool claiming otherwise is describing a different product.
- Ask the provider. For a Google file, Gemini checks for SynthID and reads Content Credentials if present, and reports which parts of a video carry the signal. Nobody knows Google’s output better than Google’s own model, and asking costs nothing.
- Check the file’s history. This is the step people skip, and it is the most common reason a result disappoints. A screenshot, a re-save, a crop, or an upload through a messaging or social app re-encodes the file. The badge will still be visible, so its presence tells you nothing about whether the pixels underneath survived. That failure looks like a broken tool rather than a damaged input.
That third point is worth internalising. The visible mark is present either way, so it carries no information about recoverability. Only the original export does.
Why vendors use two systems at all
The pairing looks redundant and is not, because the two markers are aimed at different problems. The visible badge answers a human glancing at an image. SynthID answers software, at scale, without changing how the image looks or how large the file is. Because they do different jobs, removing one does nothing for the other, which is exactly why a tool can credibly offer the first and never the second.
It also explains the tone of how vendors write about the subject. Google describes SynthID as a protection measure, and its Terms of Service prohibit both removing its branding and bypassing protective measures. Those are separate obligations, and satisfying one has never satisfied the other.
What this cannot do
- Remove SynthID. Not attempted, and no instructions given. Why →
- Handle Sora. There is no Sora page here because the tool does not process Sora output. Claiming otherwise would be a functional claim the product cannot back.
- Restore a damaged file. If the input was re-encoded, the detail is already gone and no method recovers it.
- Grant permission. See the Google Terms of Service.
Frequently asked
Can you remove an AI watermark?
The visible overlay, yes. An invisible provenance signal, no. Which one you have determines the answer, and the model name is a clue but not proof — vendors change defaults, so check the file rather than assuming.
Is it free?
Yes, with no account, no credits and no upload. The text humanizer and prompt optimizer on this site run the same way.
Will the result be blurry?
Not where the method applies, and not outside the detected region. Blur appears when a tool guesses at missing texture, or when the input was already re-encoded. See the full explanation.