SynthID bypass: screenshots, crops and compression
What Google’s documentation supports about screenshots, crops and compression—and why a changed file or a negative detector result is not a universal bypass.
Screenshots, crops and compression are not guaranteed SynthID bypasses. Google designs the watermark to survive common edits, while acknowledging that some altered images may no longer produce a detection. Those are different claims.
A different file is not evidence of a removed watermark
A ‘SynthID bypass’ claim needs a specific result behind it. Did an edit remove a visible logo, change a file’s metadata, or change a SynthID detector response? A before-and-after image can show the first kind of change. It cannot show an invisible watermark result on its own.
Google DeepMind describes SynthID as embedded directly in generated media and designed to withstand common changes including cropping and lossy compression. An edit may produce a new filename, dimensions or byte size while leaving a detectable watermark. Treat the file change and the watermark check as separate observations.
Does screenshotting remove SynthID?
Do not assume so. Google’s Gemini verification instructions explicitly discuss checking screenshots. They recommend tightly framing the image and avoiding a collage of different images. That guidance would make little sense if taking a screenshot automatically made SynthID verification impossible.
This supports a narrow conclusion: a screenshot can still be a meaningful verification input. It does not establish that every screenshot will retain a detectable watermark. Display size, scaling and what was captured can differ, so a result belongs to the particular screenshot that was checked.
Keep the original download if you have it. A screenshot is a new representation of what appeared on screen, and any detail lost in that representation should be considered separately from the watermark question.
What cropping and compression actually demonstrate
Google names both cropping and lossy compression among the changes SynthID is designed to resist. There is no support in that statement for a universal crop size, JPEG setting or number of exports that removes it. ‘Designed to resist’ also should not be rewritten as ‘impossible to defeat’.
A useful review separates the edit’s visible consequence from the evidence needed for a watermark claim.
| Change | What you can observe | What it does not establish |
|---|---|---|
| Crop the picture | The composition or visible corner changes. | That an invisible SynthID watermark is absent. |
| Compress or resize | The dimensions, file size or visible detail may change. | That a smaller file cannot carry a detectable watermark. |
| Turn off Gemini’s visible watermark | The exported image lacks the visible mark controlled by that setting. | That Google has disabled SynthID or Content Credentials. |
Source: Google: visible watermark settings leave SynthID enabled
How much does a negative result prove?
Google says common edits usually leave the watermark in place, but many alterations can make it undetectable. Record a negative response as a non-detection for that file and check. It is not proof of human authorship, a complete history of the image, or a prediction of another system’s response.
Why one example is not a benchmark
The SynthID-Image research paper evaluates detection under defined transformations and distinguishes true positives from false positives. That matters when reading a claimed success rate: results depend on which images, edits and detector conditions were tested. A single edited example cannot establish a general result across all images.
Ask what the demonstration leaves out. Were failed examples included? Was the starting image detectable? Was the final image still useful? A confident percentage without a stated sample and checking method gives a reader little basis for comparing tools or making a publishing decision.
Source: Google researchers: SynthID-Image robustness evaluation
What a useful demonstration should disclose
Before relying on a screenshot of a detector response, look for a record that connects the original, the actual output and the exact check. The following is an editorial checklist for assessing a claim, not a published benchmark or a recipe for making an image undetectable.
- Source: which authorised image was used, and whether its original produced a SynthID detection.
- Change: what happened between source and result, with each output kept separately.
- Check: the named verification service, date and complete response for the actual output file.
- Scope: whether the response concerns SynthID, Content Credentials or a generic AI estimate.
- Quality: whether important detail, text, colour and framing remain usable.
- Conclusion: a description of the observed example, without converting it into an all-image guarantee.
Source: Our guide to checking a supported image before and after processing
Choose the next step from the actual requirement
For supported Google AI images, Spoofzy offers a dedicated SynthID Removal workflow in Telegram. Its current scope is images, not every media type supported by Google’s broader SynthID technology. Check the product page for accepted inputs and limits, then review the returned file and any verification result separately.
If the job concerns GPS or camera details, use the metadata workflow instead. If it concerns a visible logo, that is a separate visible-editing task. None of these operations establishes ownership or removes a publisher’s disclosure requirements. This article reports the limits of the available evidence; it does not claim a successful test of Spoofzy or any universal bypass.
Start with one file worth preparing.
Choose the tool for the job, check its current limits and review the result in Telegram.
See the SynthID image workflow