Forensic Cleaner by Two Tensors
A detector score is not a verdict.
The case for AI image detector bypass, and the limits of it. A detector reports a probability. Automated moderation often treats it as a fact, and the person who owns the image absorbs the difference.
How an AI image detector actually decides.
It does not read a label off your file. It runs a classifier, returns a number, and someone else picks where that number becomes an accusation.
A score is an estimate, not a finding
An AI image detector returns a value for how much the pixels resemble what its classifier was trained to call synthetic. That is a statistical estimate, produced by a private model, read against a threshold set by whoever deployed it. Move the threshold and the same image changes category without a single pixel changing.
Classifiers also learn the generators that existed when their training set was built, so they generalize unevenly to newer models and to ordinary photo editing. A camera file that has been through a denoiser, an upscaler, and a few rounds of re-encoding has been resampled enough for its statistics to drift toward whatever the classifier was told to look for.
False positives are a normal output
A detector that performs well on a benchmark performs well on images that resemble its training data. Real uploads do not: screenshots, re-uploads, phone exports, scans, film grain, heavy retouching. Some detectors respond so strongly to resolution and JPEG compression that they score ordinary photographs as AI, which is why downscaling to 1024px sometimes flips their answer without changing what the image is.
Even a low error rate produces a large number of wrong calls once a service is scanning millions of files a day. Nobody is told they were one of them.
The score rarely stays a score.
By the time a number reaches you it has usually already been applied to something.
- It is applied before you see it Reach drops, a listing is held, an upload is refused, a label appears under the image. The score is the reason, and the reason is not shown.
- There is often no way to answer it Where an appeal exists it tends to run back through the same pipeline, with no field for explaining how the image was made or what rights you hold in it.
- The model moves under you Detector vendors retrain and reweight on their own schedule. A file that scored clean last quarter can score differently this quarter with nothing about the file having changed.
Owning an image should mean having a say in how it is classified.
The narrow version of the argument
There is a maximalist version of this that we are not making. Detectors are not useless, provenance research is not a conspiracy, and a platform is entitled to run whatever classifier it likes on content it hosts and pays to serve.
The narrow version is this. If you made an image, commissioned it, or hold a license to publish it, you have a reasonable interest in how automated systems classify it, and that interest does not evaporate because the classifying was done by software instead of a person. Reducing the signals a detector keys on is a technical answer to a technical claim about your file.
What the tool does about it
Forensic Cleaner reduces detector-facing signals while keeping the output close to the source image. In internal testing it reached about a 90% clean rate across three popular AI image detectors. That number is ours, measured on our own images, and it is not a promise about your image or about any detector in particular.
The workflow is deliberately unglamorous. Validation checks dimensions and safety, then an AI slop score estimates how hard the file will be to clean. Strong mode reduces signals aggressively. Quality mode, on paid plans, is gentler and leaves less visible trace of the cleaning itself. SynthID watermark removal is a separate optional stage. You can run it on an image you own and check the output against whatever detector you care about.
Where this stops.
Forensic Cleaner is for images you own or hold a valid license to process. Everything above assumes that condition.
Impersonation, forged identity or financial documents, intimate imagery of people who did not consent, any sexual content involving a minor, and anything unlawful where you are: prohibited under the terms, and grounds for termination. The complete list is in the Safety Guidelines. What gets stored, and for how long, is in the Privacy Policy.
The product also makes no claim about outcomes. It reduces detector-facing signals. It does not make a file undetectable, and it has no influence over how a platform ranks, labels, or moderates you.
FAQ
Questions about the case for bypass.
Is an AI image detector bypass a legitimate thing to use?
On images you own or are licensed to process, it is an ordinary operation on your own file, closer to re-encoding or resizing than to anything exotic. Legitimacy is decided by the purpose, not the processing. Deception, impersonation, imagery of people who did not consent, and unlawful use stay off limits regardless of who holds the rights.
Why would a real photograph be scored as AI?
Because detector scores react to things that merely correlate with generation: compression history, resampling, denoising, upscaling, low resolution. A phone photo that has been exported and re-uploaded a few times carries all of those. On detectors that are sensitive to resolution, the 1024px option sometimes changes the result.
Does this guarantee my image will pass a detector?
No. Internal testing measured about a 90% clean rate across three popular AI image detectors. Individual results vary between images and between detectors. The product reduces detector-facing signals and does not promise invisibility.
Does the workflow remove image metadata as well?
It does, as a side effect. The validation step converts the upload to lossless webp, and embedded EXIF does not survive that conversion. This falls out of the pipeline rather than being a dedicated metadata tool, so do not treat it as a feature to rely on.
Is watermark removal the same thing as detector bypass?
No. SynthID is a watermark deliberately embedded by the generator, so removing it is a different problem from lowering a statistical score. It runs as a separate optional stage, and it is not reliable on every image.
Who is this page actually for?
People who hold the rights to the files they process: photographers whose edited work gets scored as synthetic, designers publishing licensed assets, teams pushing images through platforms that moderate automatically. If the file is not yours and you have no permission to process it, none of this applies.
For the settings and the order they run in, read How to Remove AI Detection.
Related pages.
Guides
- AI Detection RemoverThe cleaner itself, and what it changes in a file.
- Bypass AI Image DetectionValidation, strength modes, and export.
- How to Get Rid of AIWhat signal reduction can and cannot do.
- Remove AI DetectionClean a single image in the browser.
- SynthID RemoverThe separate watermark stage, and its limits.