AI detection & provenance
StegoPixel reads seven independent traces a file carries — what its metadata admits, what its pixels give away, what its watermark decodes to — and fuses them into one verdict: Natural, Inconclusive or AI-generated. When the evidence is thin, it says Inconclusive. Nothing is uploaded.
The quietest evidence is written down: the names a generator leaves in EXIF and XMP, the way a camera's story contradicts itself, the odd emptiness where capture data should be. None of these carries a verdict alone — they corroborate.
A lookup table of roughly 30 known generators and editing tools, matched against the EXIF and XMP fields they tend to leave behind — software strings, prompt fields, workflow markers — each with its own weighting. It also reads the IPTC digitalSourceType field, the standard machine-readable declaration of synthetic media.
Reads the EXIF block as a story and asks whether it contradicts itself: a Make/Model brand mismatch, invalid or future dates, a modification date before the capture date, an all-zero GPS position. A camera identity with no capture parameters is noted more gently, because aggressive scrubbing looks much the same — and one contradiction never decides anything.
JPEG, WebP and TIFF files normally carry capture metadata; this check notices when the block is completely empty. It is deliberately the weakest signal — messaging apps and privacy scrubbers strip metadata as routine — so it raises Inconclusive at most, never an AI verdict by itself. PNG and other export formats are exempt: they never carried that metadata to begin with.
All three signals sit in the AI & provenance workspace alongside the pixel measurements; the detection questions that come up most are answered in the FAQ.
Erase the metadata and the image still carries its history in its own statistics. Signal 04 blends five closed-form measurements, by fixed weight, into one pixel score — and it keeps working on a file whose metadata was stripped before you ever saw it.
| Signal | Weight | Measures |
|---|---|---|
| DCT coefficient statistics | 0.30 | Skewness and kurtosis of the image's DCT AC coefficients. |
| Noise uniformity | 0.20 | How evenly noise is distributed across the frame. |
| CFA (demosaic) trace | 0.20 | The colour-filter-array pattern a real sensor leaves behind. |
| Spectral artefacts | 0.20 | Frequency-domain structure measured across the image. |
| Shot-noise correlation | 0.10 | Correlation within the photon-noise component. |
An offline frequency-domain scan for repeating watermark carriers — spectral peakiness and cross-channel phase coherence, compared against a small set of hand-authored carrier profiles. It is a heuristic pattern match, not a read of any vendor's production watermark key, and it takes two individually convincing carriers before it will assert AI.
The original Stable Diffusion pipeline embedded a fixed invisible watermark — the 136-bit payload StableDiffusionV1 — into the chroma channels of everything it generated, and this decoder reads that publicly documented bit pattern back out of the wavelet domain, block by block, by majority vote. Bit agreement of 90% or better is a high-reliability AI signal; 75–90% only argues Inconclusive; below that, a miss proves nothing, because resizing or cropping destroys the watermark.
The watermark panel sits among the forensic analysis instruments; the same detection also appears among the workflow nodes.
Signal 07 is a different kind of evidence. Many files arrive with Content Credentials attached — a CBOR/JUMBF manifest recording who claims to have created the image, with what, and from which ingredients. StegoPixel finds the manifest wherever it sits in the file, decodes it, and reports what it contains: the claim generator, title, format, instance identity, issuer details, assertions and the declared digitalSourceType.
When that declaration reads trainedAlgorithmicMedia, the file itself is declaring synthetic origin — a strong signal that costs the authenticity score 25 points, and 10 more when every manifest present is unsigned. This is reading, not verification: no signature and no certificate is validated, and this page will not imply otherwise.
Content Credentials record who asserts what about an image's origin. StegoPixel surfaces that claim exactly as found — it does not validate signatures or hashes, and it cannot tell you whether the photograph depicts reality.
The same reading runs headless as the C2PA_Check node and the C2PA Provenance Check preset in the workflow library; what a manifest can and cannot prove is covered in the FAQ.
Seven families of evidence, each contributing what it can, fused into one conservative answer. The rules are simple, and deliberately hard to fool.
The same verdict runs headless: the AI Detection node in four modes, plus presets including AI-Generated Detection, Full Authenticity Report and Provenance & AI Audit — see the workflow presets.
Alongside the verdict, the Home module keeps an authenticity score — a weighted summary of everything measured for the loaded image: high, medium and low-risk inconsistencies, hidden payloads, AI indicators, provenance nobody signed. It resolves to one of three bands, with the breakdown visible so you can see which finding moved the number.
| Score | Band |
|---|---|
| 80–100 | Likely Authentic |
| 50–79 | Uncertain |
| Below 50 | Suspicious |
The bands are stated openly because a number you cannot interpret is worth very little.
See also the forensic recipes and reports.
A detector that never admits its edges is worth less than one that does. Read the boundaries before you rely on anything above.
Every offline detector produces both false positives and false negatives. Use the contributing signals rather than the headline label, and never present a score as a determination about a specific person or image.
You read an image's claims before you repeat them — what its metadata admits, what its pixels give away, what its provenance says. The verdict, the working and the doubts all sit in front of you.
You want a provenance reading that is honest about being a reading — and a report someone else can follow line by line.
You want to know what a picture carries, and what survives once a platform re-encodes it.
You want the signals, the score breakdown and the file digests together, in one HTML or PDF report.
Questions before you commit are collected in the FAQ; when you are ready, see how to get it safely.
It combines seven offline signals — generator signatures in metadata, a plausibility audit, a stripped-metadata check, pixel statistics, a carrier-signature scan, the Stable Diffusion watermark decode and C2PA parsing — and resolves them into one of three tiers: Natural, Inconclusive or AI-generated. To be clear about what is under the bonnet: these are hand-written statistical heuristics, not a trained model. There is no model file and no labelled dataset, which is exactly why the app reports confidence tiers rather than a yes/no answer.
No, and you should be suspicious of any offline tool that claims it can. Every detector of this kind produces both false positives and false negatives, and metadata can be forged, stripped or added innocently by an editor. Use the contributing signals rather than the headline label, and do not present a score as a determination about a specific person or image.
The verdict is Inconclusive, which is a real outcome rather than a failure. Asserting AI takes at least one high-confidence signal or two medium ones; confidence on heuristic evidence is capped at 90%; and a check that could not complete marks the verdict incomplete instead of clean. Every verdict also lists which checks ran, were skipped or failed, so the reasoning stays visible.
Yes. There is no HTTP client and no telemetry anywhere in the application — detection runs exactly where your images already live, and disconnecting the network changes nothing about how the app behaves. The only outbound links it can produce are optional map references written into a geo-tag report, and those are followed only if you click them.