How to make an AI image undetectable
An AI image gets flagged for three reasons: the provenance block the generator writes into the file, the absence of camera sensor noise, and frequency patterns left by the model. To make it pass, strip the metadata, add realistic sensor noise and lens behaviour, smooth the frequency spectrum, then re-encode it as a camera JPEG with camera EXIF. PassReal does all four in one upload, and refunds the payment if a public detector still flags the image.
Last updated: 2026-10-08
Why AI images get flagged
A detector does not look at an image the way a person does. It reads layers that the eye skips, and an AI export fails on several of them at once.
The first layer is the file itself. Generators such as Gemini, ChatGPT and Adobe Firefly write a provenance record into the export: a C2PA manifest, an IPTC field that says the image came from a trained model, and often an XMP entry with the tool name. On one Gemini export we counted 174 such markers. Any platform that reads metadata sees the label before it looks at a single pixel.
The second layer is noise. A photo taken with a phone carries grain from the sensor, and that grain follows rules: it is stronger in the shadows, it differs between the red, green and blue channels, and the lens adds colour fringes that grow toward the corners. A generated image is too clean, or clean in the wrong way.
The third layer is frequency. When an image is moved into the Fourier domain, the output of a diffusion model shows regular peaks in the high frequencies. Classifiers trained on large sets of real and generated images learn to spot them.
The last layer is encoding. A camera saves a JPEG with its own quantization tables and writes EXIF fields that agree with each other. A PNG straight out of a generator has none of that.
The manual method
You can clean each layer yourself with free tools. It takes some care, and it is worth doing on a copy.
- Metadata: ExifTool reads and deletes almost every family. Delete, then read the file again group by group. A single forgotten PNG text chunk can still hold the full prompt.
- Noise: an editor such as GIMP or Photoshop can add grain, but uniform grain looks fake. It should follow brightness and differ per channel.
- Frequencies: a Fourier low-pass is available in scientific tools such as Python with NumPy. Too strong, and the image goes soft.
- Encoding: export a JPEG at a realistic quality, then write coherent EXIF fields for one real phone model.
Each step is simple on its own. Getting all four right on every file, without visible damage, is where it goes wrong.
The one-upload method
PassReal runs the same four steps as a single pipeline. It strips every metadata family, simulates a camera sensor with brightness-dependent grain and radial lens fringing, smooths the high frequencies, and re-encodes the result as a camera JPEG with matching EXIF. The marker count goes from 174 to 0 on a Gemini export, and every image comes back with its SSIM and PSNR against your source, so the quality cost is measured instead of guessed.
The money-back guarantee
If a publicly available AI image detector classifies an image you paid PassReal to process as AI-generated, email us within 14 days with the image and a screenshot of the result. We refund the payment that covered it, the single image credit or the Pro month. The terms are in section 7 of our Terms of Service. Free images are not covered, since nothing was paid for them.
How to check the result yourself
Do not take any tool at its word, ours included. Upload the processed file to two or three detectors and keep a screenshot of each verdict. Then open it in a metadata viewer and confirm that no C2PA, IPTC or XMP entry names a generator. If something is left, you know which layer to fix.
For the detail of each layer, read how to remove AI metadata from an image and how to add camera noise to an AI image.
Questions
Which detectors does this work against?
The method targets the signals that public image detectors read: metadata, noise statistics and frequency patterns. With PassReal, if a public AI image detector flags an image you paid to process, email us within 14 days with the result and we refund that payment in full.
Does making an image undetectable lower its quality?
Each step is applied at low strength. PassReal measures SSIM and PSNR between your source and the result for every image, so you can see the change in numbers rather than guess.
Is removing the metadata enough on its own?
No. Metadata is the easiest signal to read, so it goes first, but a detector that analyses pixels ignores it. Noise, frequencies and encoding have to be handled as well.
Can a Google SynthID watermark be removed?
SynthID sits inside the pixel values and Google says it was built to survive edits and compression. We do not claim to remove it. The guarantee covers public detectors, which is what most platforms and buyers use.
Do I still have to disclose AI use?
Platform rules and local law can require you to say an image is AI-generated. Processing the file does not change that duty, so check the rules of the place where you publish.
Limits
Naturalization operates on the pixels and on the file, not on the content of the image. A subject that is implausible - six fingers, inconsistent reflections, text that does not read - stays implausible after processing, and a human reviewer will notice it. PassReal changes what a statistical classifier measures, not what a person sees.
Results vary by generator, by subject and by detector, and detectors are retrained. No pass rate is published on this page because none has been measured in a way that would still hold next month.
Three images, no card, no expiry.