Is this image real?

rymnis.ai runs your image through a pipeline of independent signals — metadata, cryptographic provenance, and ML classifiers — and tells you exactly which ones fired and why.

cloud_upload

Drop an image here, or browse

JPEG · PNG · WebP · TIFF — up to 32 MB

videocam Video detection is coming.

We're building it. Want to know when it's ready?

article Text detection is coming.

We're building it. Want to know when it's ready?

mic Audio detection is coming.

We're building it. Want to know when it's ready?

How it works

No single technique is a silver bullet. We ensemble multiple independent signals and show you exactly which ones fired.

fingerprint
Provenance & metadata

C2PA cryptographic manifests, SynthID watermarks, and EXIF software tags provide high-confidence signals that don't depend on ML classifiers.

psychology
ML classifiers

AIDE — a deep neural network combining DCT frequency analysis and a frozen ConvNeXt-XXL backbone — scores the image for synthetic origin, generalising across many generator families.

account_tree
Ensemble verdict

Signals are weighted by reliability. High-confidence cryptographic signals override weaker statistical ones.

Detection signals

Every signal included in the current pipeline.

EXIF / Metadata fast

Reads all EXIF fields via ExifTool. Matches the Software tag against known AI generators (Stable Diffusion, Midjourney, DALL-E, etc.) and checks for camera manufacturer signatures.

C2PA Manifest fast

Verifies the cryptographic Content Provenance (C2PA) manifest if present. A valid manifest from a known AI generator is a high-confidence signal. Supports c2patool v0.26+ format.

SynthID fast

Detects images generated by Gemini and other Google AI products. Google does not provide a public SynthID API, so detection is inferred from C2PA manifest assertions embedded by Google's tools.

AIDE Detector ML

AIDE (AI Image Detection via DCT and ConvNeXt) runs on a GPU and outputs a probability score for AI generation. It combines DCT frequency-domain patches with a frozen ConvNeXt-XXL visual backbone, achieving 92.77% on AIGCDetectBenchmark and 86.88% on GenImage.

A note on accuracy
92.77% AIDE on AIGCDetectBenchmark
86.88% AIDE on GenImage

Detection is a moving target. Modern high-quality generators have significantly closed the gap. We report which signals fired and at what confidence rather than a black-box verdict, so you can weigh the evidence yourself. A high-confidence cryptographic signal (C2PA, SynthID) is reliable; an inconclusive ML result means the classifier couldn't decide — not that the image is real.