Forensic Cleaner by Two Tensors

SynthID Remover for AI Images

Google DeepMind embeds an invisible watermark in Gemini and Imagen output. Forensic Cleaner treats removing it as its own stage.

What SynthID actually is

Google DeepMind describes SynthID as a way to embed an imperceptible digital watermark directly into generated content while leaving the quality of that content alone. In an image it lives in the pixels, spread across the frame rather than stamped into one corner. There is nothing to find by looking, at any zoom level, on any monitor.

What makes it worth its own tool is durability. It is built to outlast the edits people assume will destroy a watermark: a crop, a resize, a screenshot, a fresh save in a different format. Google tooling can then read the file later and report that it came from a Google model, and that answer follows the image into places you never uploaded it to.

Which images carry it

If you generated or edited an image with Gemini or Imagen, assume the output carries it. Google applies SynthID across its generative products, and DeepMind describes the same approach for audio, text, and video as well. Forensic Cleaner works on images only, so PNG, JPG, and WEBP are the formats that matter here.

How to tell it apart

  • Invisible at any zoom, in any editor
  • Carried in the pixels, not in file metadata like C2PA
  • Deliberately embedded, not a detector's statistical guess
  • Built to survive crops, resizes, and screenshots

Watermark removal and detector cleaning are two different jobs.

SynthID removal is an optional stage that runs in front of the normal cleaner, with its own switch and its own strength control. A deliberate embedded signal and a detector's statistical guess are not the same problem.

  • SynthID Removal Targets the embedded watermark signal in the pixels. Enable it before you start the job and pick one of four strengths. This stage runs first.
  • Detector Bypass Targets the signals third-party AI detectors score. Strong and Quality modes belong to this stage, not to the watermark stage. Quality mode needs a paid plan.
  • Both, in one job Leave Detector Bypass on and the watermark output feeds straight into the cleaner, so you download one final image. Turn it off for watermark-only output.

SynthID-only processing costs 3 credits. Keeping Detector Bypass enabled adds those 3 credits on top of the usual cleaning cost for the image. Allowances per plan are on the pricing page.

Four strength settings. Start at medium.

Strength decides how hard the stage pushes at the watermark. Pushing harder also takes more out of the image, so this is a trade you make deliberately, not a quality dial you turn all the way up.

low

The gentlest pass. Worth trying first when the frame has small or distant faces, fine hair, or small text, since those details break first under heavier processing.

medium

The setting to start on. Run it, check the output at full size, and move up only if the watermark is still being reported. Most images never need more.

high

An escalation, not a default. Reach for it when medium leaves the signal readable and the subject is forgiving enough to absorb a harder pass.

x_high

The most aggressive setting available. Best odds on a stubborn image, and the largest visible cost. Treat it as the last thing you try, not the first.

There is no prize for starting high. Escalate one step at a time and compare each output against the source at full size, because the detail a heavier pass costs you is gone from that result for good.

Chaining the two stages

Open the Forensic Cleaner, upload a PNG, JPG, or WEBP you have the rights to process, and switch SynthID Removal on before the job starts. With Detector Bypass left on, one job runs both passes in order and gives you a single download. With it off, you get the watermark result by itself, which is what you want when the detector score is not your problem.

Resolution helps more than you would expect

If an image still comes back watermarked, try the 1024px downscale option before you raise the strength. Downscaling improves SynthID removal reliability, and it often gets there with less visible cost than a harder pass at full size. The free tier caps output at 768px and paid plans unlock 1024px.

Automating it

Pro and Scale include API access. Create a job against POST /api/v1/bypasses with remove_synthid_watermark set to true. detector_bypass defaults to true, so set it to false when you want the watermark stage alone. Poll the job, then download the output once the stages you selected have finished. The API docs cover uploads, idempotency keys, and job states.

If the watermark is not your real problem

Plenty of people land here after a detector flagged an image that never touched a Google model. There is no watermark to remove in that case, and the other stage is the one you want: AI Detection Remover covers it. If you are still working out what any of this can honestly do to a picture, How to Get Rid of AI is the plain version, and why creators clean their own work covers the reasoning.

Where this stage stops working.

Removal is not guaranteed on every image. Some frames come back clean at medium and others are still reported at x_high. When the watermark signal is deeply embedded there may be no setting that clears it while leaving an image you would publish, and pushing past that point trades picture quality for nothing.

When repeated attempts fail, the honest answer is usually a different source image rather than another rerun. Regenerate the frame, or start from one that was not produced by a Google model in the first place.

Nothing here is a promise about a platform. Forensic Cleaner reduces detector-facing signals. It does not guarantee that any specific detector, provenance dashboard, or moderation queue will accept the result. Use it on images you own or hold a license to process. Impersonation, fraud, non-consensual imagery, and other illegal uses are out of bounds.

FAQ

SynthID removal questions.

Can I see SynthID in an image myself?

No. It is embedded below the threshold of vision, so zooming in, pushing the levels, or opening the file in an editor will not reveal it. Detection tooling reports it. Your eyes cannot.

Does cropping, screenshotting, or re-saving strip the watermark?

Those are the edits the watermark is designed to survive. A screenshot re-encodes the same pixels, and a crop keeps most of them, so the signal usually travels with the new file. That is the reason a dedicated removal stage exists at all.

Do I have to run detector cleaning as well?

No. Turn Detector Bypass off and the job runs the watermark stage on its own. Leave it on and the same job chains both, so the image is watermark-processed first and then cleaned for detector-facing signals.

Is it worth starting at x_high to save a rerun?

No. Higher settings work harder at the watermark and take more out of the image, and that detail does not come back in the output you download. Start at medium, judge the result, and step up one level at a time.

Why did my result come back distorted?

The strength was probably higher than the image could carry. Drop one level and run it again. Faces that occupy a small part of the frame, fine hair, and small text show the cost before anything else does.

Does this remove C2PA or Content Credentials?

No, and they are a different kind of problem. C2PA provenance travels as metadata attached to the file rather than as a signal inside the pixels. Separately, uploads are converted to lossless WEBP during validation, and that conversion drops the file's existing EXIF as a side effect of the pipeline.

SynthID also covers audio, text, and video. Can this handle those?

Not today. Google DeepMind describes the same watermarking approach across several media types, but Forensic Cleaner is an image pipeline and accepts PNG, JPG, and WEBP only.

For what happens to an uploaded file, read the Privacy Policy. For what you may process, read the Safety Guidelines.