Get Started Free 1 image, no card

Your export says
"made by AI".
Ours doesn't.

PassReal strips the provenance block a generator writes into your file, then rebuilds it with a real camera profile. 174 markers in, 0 markers out.

Get Started Free See how it works
Tested against
Hive Illuminarty AI or Not GPTZero
See the difference

From declared to stripped

Drag the slider. Same frame, same visual quality. Only the sensor signature and the metadata differ.

The same café scene after running through the PassReal pipeline No provenance blockNo provenance After PassReal
An AI-generated café scene before processing C2PA provenance blockC2PA tagged Original AI Image

A real AI-generated image, processed by the live PassReal pipeline. Both halves are the same frame - only the sensor characteristics differ.

The file itself

Your image stops declaring itself.

Every generator signs its output. Here is a real Gemini export, and the same file after the pass.

174 → 0 provenance markers left in the file
Declares AI Gemini export · PNG · 1.5 MB
PNG chunkcaBX · 15.1 kB
Manifestc2pa.claim.v2, signed
Signed byGoogle Media Processing
Cert chainGoogle C2PA Root CA G3
TimestampGoogle TSA, countersigned
SourceTypetrainedAlgorithmicMedia
Makeabsent
Modelabsent
Exposureabsent
Lensabsent
Declares nothing PassReal output · JPEG · 131 kB
Provenancenone
MakeApple
ModeliPhone 15 Pro
Software17.4.1
Exposure1/60 s · f/1.78
ISO713
Focal6.765 mm · 24 mm eq.
Lensback triple camera f/1.78
Shot at2026:07:01 10:18:20
QuantTablematches the make
stripsimulatesign

Drop a file in, get it back. The readout above is an actual before and after - not an illustration.

The chain

Four stages run. Two are switched off.

Nothing here is guesswork. The two stages that made images easier to spot are still in the code, switched off, with the numbers that did it.

Erase the origin Every AI tool hides a signed certificate inside the file saying it made the picture. That is what platforms read. It goes. 174 → 0 C2PA manifest · IPTC source type · XMP · text chunks
Extra filtering We tried a second cleaning pass. It fooled no detector any better and it softened the picture, so it stays off. OFF per-block DCT · no AUC gain outside the noise · costs 1 dB PSNR
Put the camera back Real photos have grain, and colours drift slightly toward the edges of a lens. Generated images are too clean. We add that back. grain · lens drift Bayer CFA and demosaic · ISO grain 0.018 · lateral CA 1.25 px at the corners
Smooth the giveaway Generators build a picture up in steps, and leave a faint regular pattern behind. Detectors look for it. This takes the edge off it. pattern softened low-pass on upsampling residue at 0.88 × Nyquist
Redraw with AI We tried rebuilding the picture with Stable Diffusion. It came out easier to detect, not harder, and visibly damaged. Off since August. OFF SDXL · 79.2% flagged against 32.5% · SSIM 0.736 against 0.918
Sign it as a photo The file leaves with a camera make, a model, a shutter speed and a lens - and compressed the way that camera compresses. looks identical EXIF from a real device profile · matching quantization table · SSIM 0.87-0.90

Batch Up to 20 images in one go. Upload, walk away, come back.

API Access on request - write to us and we set it up.

Pricing

Start free. Scale when ready.

No credit card required. Upgrade anytime.

Free

$0

Perfect for trying it out

  • 1 image now, 2 more once your email is confirmed
  • Standard quality
  • JPEG output
  • Web interface
Get started
Most Popular

Pro

$19 /month

For creators and professionals

  • 100 images per month
  • Maximum quality pipeline
  • Batch processing (20 at once)
  • Priority queue
  • PNG + JPEG output
  • API access
Get Pro

Enterprise

Custom

For teams and high volume

  • Unlimited images
  • Dedicated API endpoint
  • Custom SLA
  • Priority support
  • Volume discounts
  • Custom integration
Contact sales

Just need one image?

Pay per image, no subscription. Credits never expire and stack on top of any plan.

$21 image $1210 images
FAQ

Common questions

Our pipeline maintains a SSIM score above 0.85 compared to the original - meaning the visual difference is imperceptible to the human eye. The output is a high-quality JPEG (93-96 quality) with realistic camera characteristics.
An AI image humanizer takes a generated image and makes it behave like camera output: generation metadata is stripped, sensor grain and lens aberration are applied, frequency-domain residue is smoothed, and camera EXIF is written with the matching JPEG quantization table. PassReal is an AI image humanizer for PNG, JPEG and WebP files up to 8192 px.
No pass rate is published, because any single figure would be out of date the moment a detector is retrained. What is measured and reported per image is fidelity: SSIM and PSNR against your source. Test your own files before and after - most detectors offer a free tier.
Partly, and it depends on what the platform checked. Most rejections start with a script reading the file: PNG text chunks holding your prompt, XMP blocks, or a C2PA manifest declaring the image machine-generated. PassReal removes all three and writes coherent camera metadata instead. It does not change whether the artwork itself reads as machine-made to a human reviewer.
Instagram's automatic label is triggered mainly by industry provenance metadata - C2PA manifests and IPTC digital-source-type fields that generation tools write into the file. Stripping that metadata and writing camera EXIF removes the signal the label reads. Pixel watermarks such as SynthID are a separate layer and are not claimed to be removed.
Yes, and it is the most common use. Upload the finished square at its largest size - distributors want 3000×3000, and processing a small export then upscaling undoes the sensor grain. Run it last, after any lettering or colour grade, so the whole frame carries the same treatment.
Processing typically takes 20-60 seconds depending on image size and server load. You'll see real-time progress updates as your image passes through each pipeline stage.
Yes. Images are processed and then automatically deleted from our servers. Source images are removed immediately after processing (unless you opt to retain them), and results expire after 7 days. We never use your images for training.
PassReal is an image processing tool, similar to photo editors or filters. How you use the output is your responsibility. We recommend using PassReal ethically and in compliance with applicable laws and platform terms of service.

Strip the provenance from your next export

Get Started Free