Forensic Cleaner by Two Tensors

How to Get Rid of AI on an Image

At least four different problems hide behind that phrase. This page works out which one you have, what Forensic Cleaner does about it, and where it stops.

Open Forensic Cleaner View pricing

Four problems, one phrase.

People searching this term are rarely asking the same question. Work out which of these you are looking at before you pick a tool, because only some of them have a technical fix.

Detector signals

The statistical fingerprint that third-party AI image detectors read and score. This is the part Forensic Cleaner is built to reduce.

SynthID

An imperceptible watermark embedded by Google models. You cannot see it, and it is not stored in the file's metadata. It needs its own stage.

File metadata

The EXIF fields carried inside the file: camera, software, sometimes location. The easiest of the four to get rid of, and the least interesting.

The look

Some images read as generated to a person within a second. No cleaner changes that, and no detector needs to be involved for it to be a problem.

What this product addresses, and how well.

Detector signals: the main job

Forensic Cleaner is a detector-signal reducer. It rewrites the traces that AI image detectors key on while holding the output close to the source, because an image that passes but no longer looks like your image is not a result. Internal testing measured about a 90% clean rate across three popular AI image detectors. Treat that as a description of our own runs rather than a forecast for your file. If you want the mechanics of what those detectors actually read, the page on bypassing AI image detection goes signal by signal.

SynthID: a separate, optional stage

SynthID is not part of the normal clean. It is an extra stage you switch on, it costs extra credits, and it is worth switching on only when the image came out of a Google model. It is also less consistent than detector-signal reduction: some images clear on the first attempt and some never do. The SynthID remover page covers the settings, the strengths, and the credit cost.

Metadata: a side effect, not a feature

Validation converts every upload to lossless webp, and the EXIF block does not survive that conversion. So the file you download has lost the camera, software, and location fields the original carried. That is useful and it costs you nothing extra, but it is a by-product of the pipeline rather than a metadata tool, so it is not a reason to choose this product on its own.

The look: not a software problem

Nothing here touches the fourth problem. Six fingers, melted text on a sign, that particular plastic sheen: a clean run leaves all of it exactly where it was. If a person can call the image at a glance, fix the image.

In one line each

Reduces detector signals About a 90% clean rate across three popular AI image detectors in internal testing. Never a guarantee, and never aimed at one named service.
Optionally removes SynthID A separate stage you enable per job, with its own credit cost and its own failure rate.
Drops EXIF metadata A consequence of the lossless webp conversion during validation, not a dedicated stripper.
Leaves the picture alone Subject, framing, and anything a viewer would notice stay put. Visual similarity is the constraint the model works under.

Why the same settings give two different results.

Detectors do not agree with each other

There is no single AI detector. There is a shifting set of third-party services, each trained on a different mix of images, each with its own threshold, and each shipping model updates without telling anyone. An output can read clean on one service and score high on another the same afternoon. That is why the cleaner targets broad signal reduction instead of tuning against one named detector, and it is why a single test tells you very little. Check two or three.

Some detectors are also unreliable in ways that have nothing to do with your image. Several of them flag ordinary photographs as generated, because their score tracks resolution and JPEG compression more closely than it tracks how the picture was made. If a detector calls your camera roll synthetic, its verdict on a cleaned file is not worth much either.

Images are not equally difficult

Before anything is cleaned, the app shows an AI slop score: an estimate of how hard this particular image will be. A high score means the generated fingerprint is dense and spread widely through the frame, so more of it has to be disturbed before a detector stops seeing it. On those images the clean works harder, visible artifacts become more likely, and a second run is a realistic outcome. A low score usually clears on the first attempt at default settings.

Small or distant faces are the other common complication. Fine facial detail is where an aggressive pass shows itself first, which is why lowering the strength often fixes a result that came back smeared. The score sits in the cleaner next to your upload, before the credit is spent, so this is a decision you make with information rather than after the fact. If you have not run a job before, the step-by-step walkthrough shows where each of these controls sits.

Two limits worth stating plainly. Forensic Cleaner reduces detector-facing signals; it does not make an image undetectable, and it makes no commitment about whether a particular detector or platform will accept the output. A detector that updates next month can also score a file you cleaned today.

The only performance figure we publish is about a 90% clean rate across three popular AI image detectors, from internal testing. Your image is a sample of one and can land on either side of that.

When the honest answer is a different image.

Rerunning a difficult file is often the wrong move. These are the cases where changing the input beats changing the settings.

Signs you are fighting the image, not the tool

The slop score is high before you spend anything The estimate exists so the decision happens first. On a high score, generating or shooting a replacement usually costs less than three cleaning runs.
Strong clears it, but the result is smeared or tinted Quality mode, on paid plans, is the gentler pass and normally fixes this. If Quality is too gentle to clear the detector and Strong is too visible to use, nothing between them will fix that.
The faces are small or far from the camera Lower the strength first. If a strength low enough to protect the faces is too low to clear anything, recrop or reframe instead.
SynthID survives several attempts Change the strength and try the 1024px downscale, which improves SynthID reliability. After a few failed runs, a different image is a better use of credits than a fifth try.

No tool clears every image, and this one is no exception. The slop score is there so you can find that out before you spend credits rather than afterwards.

Straight answers about what to expect.

What actually changes when you get rid of AI on an image?

The detector-facing signal changes. Forensic Cleaner rewrites the statistical traces that AI image detectors score, and it is built to keep the visible result close to the source.

The subject, the framing, and the content do not change. If an image reads as generated to a person looking at it, cleaning it does not make it look like a photograph.

Is a cleaned image undetectable?

No, and nobody can honestly promise that. The product reduces detector-facing signals. It does not make an image invisible, and it does not commit to any named detector or platform accepting the output.

Internal testing measured about a 90% clean rate across three popular AI image detectors. That is a measurement of our own runs, not a guarantee about yours.

Why did it work on one image and fail on the next?

Two reasons, usually. Images differ in how dense the generated fingerprint is, which is what the AI slop score estimates before you spend a credit. Detectors also differ from each other and change over time without announcing it.

Testing against one detector tells you very little. Check the same output against two or three before deciding the clean failed.

Do I need the SynthID stage?

Only if the image came out of a Google model such as Gemini or Imagen. SynthID is an imperceptible watermark those models embed, and the normal clean is not aimed at it.

It is a separate optional stage with its own credit cost and its own failure rate, so switching it on by default just spends credits.

Does this strip the metadata from my file?

In practice, yes. Validation converts every upload to lossless webp, and the EXIF block does not survive that conversion, so the file you download no longer carries the camera, software, or location fields the original had.

It is a side effect of the pipeline rather than a dedicated metadata tool. If metadata is the only thing you want gone, an image editor will do it without spending a credit.

The AI slop score came back high. Should I still run it?

You can, but expect a harder job. A high score means the fingerprint is dense and spread widely through the image, so more of it has to be disturbed before a detector stops reading it. Artifacts and repeat runs are both more likely.

The score is shown before the credit is spent precisely so you can decide. On a high score, a different source image is often the cheaper answer.

How many attempts are reasonable before giving up on an image?

Try the obvious variations first: the other strength mode, and the 1024px downscale, which helps on detectors that react to resolution and on SynthID removal. That is usually two or three runs.

If none of those clear it, the image itself is the constraint. Further attempts mostly spend credits.

Can I run this on any image?

Only on images you own or hold a valid license to process. Impersonation, deception for gain, sexual imagery of anyone who did not consent, and anything otherwise unlawful are all off limits.

The Safety Guidelines list the restrictions in full. Read them before you upload something you are unsure about.

For data handling details, read the Privacy Policy. For use restrictions, read the Safety Guidelines.