Visible Gemini watermark vs SynthID: what actually changes?

Gemini-generated images can carry more than one kind of provenance signal. The small semi-transparent mark you can see near a corner is not the same technology as Google’s invisible SynthID watermark. Content Credentials are different again.

That distinction matters. Removing a visible corner logo changes visible pixels in one local area. It does not establish that an image has no SynthID signal, does not verify the image’s origin, and does not turn an AI-generated image into a human-made one.

That boundary is deliberate, and it is the boundary the local browser eraser enforces: it processes a measured visible mark on your device and leaves every other provenance layer untouched.

Three provenance layers that are easy to confuse

The word “watermark” is often used for all three layers, but they serve different purposes and behave differently when a file is edited.

1. The visible corner mark

A visible watermark is part of the rendered image a person sees. In supported Gemini outputs, the project predicts a likely logo size and anchor from the image dimensions, searches locally around that candidate, and validates the proposed restoration before accepting it.

The restoration math is based on inverse alpha compositing. If a logo color was blended over an original pixel with a known alpha value, the observed pixel can be written as:

watermarked = alpha × logo + (1 − alpha) × original

Solving for the original estimate gives:

original = (watermarked − alpha × logo) ÷ (1 − alpha)

The equation is simple; reliable use is not. The alpha profile, position, scale, edge antialiasing, and source pixels all have to match closely. That is why the engine does more than erase a rectangle or blur a corner.

2. SynthID

Google DeepMind describes SynthID as a digital watermark embedded directly into AI-generated images, audio, text, or video. For images and video, Google says the signal is imperceptible to people and is designed to survive common modifications such as cropping, filters, resizing, and lossy compression.

Google’s public verification guidance also says detection can be uncertain. Very simple content may not contain enough detail to watermark, minor edits may be too small to carry a detectable signal, and repeated modification can eventually affect detectability. A positive or negative result therefore needs to be interpreted in the context of the verifier’s own wording.

Most importantly for this tool: processing the visible logo is not a SynthID operation. The visible-logo algorithm does not invoke a SynthID detector, and its output should not be described as “untraceable,” “undetectable,” or “SynthID-free.”

3. Content Credentials

Gemini Help describes Content Credentials as a digital passport that can document a file’s origin and edit history. Content Credentials use C2PA technology and are not the same thing as either a visible logo or SynthID.

Provenance metadata can also be preserved, stripped, replaced, or stored remotely depending on the producing and editing tools. A visible-logo result alone tells you nothing conclusive about that metadata. If provenance matters to your workflow, inspect it separately and keep the original file.

What the visible-logo remover actually does

The current engine follows a conservative sequence:

1. Catalog lookup: known Gemini output dimensions provide candidate logo size and anchor priors.
2. Local candidate search: the engine looks near expected positions rather than assuming every image is identical.
3. Alpha-profile restoration: it applies a calibrated inverse compositing model to a candidate region.
4. Residual and damage checks: the candidate must pass validation before the edited pixels are accepted.
5. Fail closed: if confidence is insufficient, the safer result is to leave the image unchanged.

This boundary is intentional. A general-purpose eraser might accept a brush mask and invent replacement texture. This project instead targets a narrow, measured pattern. The tradeoff is that unsupported or transformed inputs may be skipped.

How to verify SynthID instead

Google currently directs signed-in Gemini users to upload an image, video, or audio file and ask whether it was created or edited by Google AI. Gemini can then check for SynthID and, where supported, Content Credentials.

That verification has a narrower meaning than a universal AI detector:

• If SynthID is detected, Google says all or part of the media was created or edited by Google AI.
• If SynthID is not detected, the file could still have been produced by another AI system.
• Inconclusive results can occur, including with very simple content or small edits.
• Content Credentials can offer separate origin and edit-history information when present and supported.

Use the verifier’s exact result rather than translating it into a stronger claim.

Which check do you actually need?

Open the branch that matches your goal. Each branch identifies a different layer.

1. 1Can I remove the small visible Gemini logo? — You can see a recognized corner mark

   Use the visible-logo tool on a copy of the original file. Review the result at full size and retain the original.

2. 2Was this file generated or edited by Google AI? — You need provenance verification

   Use Gemini’s SynthID verification flow. The visible-logo remover is not a provenance detector.

3. 3What edits and tools are recorded for this asset? — You need an edit history

   Inspect Content Credentials with a compatible verifier. Preserve the original file and any project records.

4. 4Can I claim the result is human-made? — The source was generated by AI

   No pixel edit changes authorship history. Disclose AI involvement wherever your audience, client, platform, or policy expects it.

A better publication workflow

For design production, the most defensible workflow is simple:

1. Keep the original Gemini output.
2. Record the prompt, generation date, and editing steps when provenance matters.
3. Process only the visible logo if you have the right to do so.
4. Review the result at 100% zoom, especially around fine lines and high-contrast edges.
5. Preserve accurate AI-generation disclosure in captions, metadata, project records, or Content Credentials as appropriate.
6. Use Gemini’s verification flow when you need to discuss SynthID.

This gives collaborators two things a “clean-looking” file cannot: an auditable source trail and accurate language about what was changed.

Frequently asked questions

Does removing the visible Gemini logo remove SynthID?

No claim like that is supported. This project does not detect or remove SynthID. Google describes SynthID as an imperceptible signal embedded into the generated content, separate from a visible logo overlay.

Can a cropped image still contain SynthID?

Google says SynthID is designed to remain detectable after common edits including cropping, although it is not foolproof and detectability can change after extensive modification.

Does no visible logo mean no AI was used?

No. A missing corner mark is not evidence of human authorship. Other Google AI surfaces, API workflows, edits, crops, or entirely different generation systems may produce files without the same visible pattern.

Is Content Credentials the same as image metadata?

Content Credentials are structured, signed provenance data rather than an ordinary descriptive EXIF field. They may document origin and edits, but availability depends on the tools and file history.

Where to go next

• Process a visible mark on your device with the local browser eraser, or use the Veo and Gemini video remover for clips.
• Follow the step-by-step browser removal walkthrough if this is your first run.
• Confirm your file is in scope with the supported sizes and formats guide.
• Settle the rights and disclosure questions first with how to remove a Gemini watermark responsibly.