How SafPix works
SafPix is a small kit of free, private photo tools — a forensic checker for spotting AI fakes and edits (plus a reverse-image-search hand-off to check whether a photo appears elsewhere), plus quick tools to clean, blur, stamp and sanity-check your own pictures. Every one of them runs entirely in your browser. Nothing is ever uploaded to a server.
One toolkit, five jobs
SafPix started as a real-vs-fake checker and grew into a small set of photo-safety tools. Each one does a single job well, and each runs 100% on your device — no upload, no account, no cost.
- Check — Is this photo real, AI-generated, edited or stolen? Drop any image for a forensic read with a plain-English verdict and visual overlays. (That's the deep dive below.)
- Clean — Strip the hidden data out of your own photos — GPS location, device, date — before you share them.
- Blur & hide — Cover faces, license plates, names or account numbers with a box, brush or emoji.
- Stamp — Watermark a verification selfie (“for verification only”, a name, the date) so a scammer can't reuse it.
- QR check — Read a QR code from a photo and see where it really goes — with a phishing warning — before you tap.
The first tool — Check — is the one with the forensic engine. Here's exactly how it reads an image.
Inside Check: the signal categories
When you drop an image into Check, a battery of independent passes runs in your browser — 18 forensic signals across the categories below, plus an in-browser AI detector. Forensic analysis is a game of weighted evidence, not a single oracle: no individual signal is bulletproof, but several independent signals pointing the same way is a pattern worth trusting. Here's what each category reveals.
Every real camera embeds a digital fingerprint into every photo it takes — the camera make, model, date, and sometimes GPS coordinates. SafPix reads this fingerprint and checks it against a database of known camera signatures and AI generator tags.
If the metadata says Midjourney or Stable Diffusion, the image is instantly flagged. If it says Apple iPhone 15 Pro with a valid date, that's strong evidence of a real capture. If the metadata is completely stripped — that's suspicious on its own.
SafPix also reads Content Credentials (C2PA) — a tamper-evident "nutrition label" that newer cameras, Adobe apps and AI tools now attach to images. When it's present and the signature is intact, it's the strongest evidence there is: it can cryptographically prove a photo came from a real camera, or that it was generated by a specific AI tool. SafPix lets a valid Content Credential override every other signal — and flags it if the signature has been broken.
When someone pastes a face onto another photo, adds an object, or composites from multiple sources, the edited region compresses differently from the rest of the image. SafPix's compression analysis detects these inconsistencies and renders them as a visual heatmap — green means consistent, red means something was changed.
Honest limitation: This signal is strongest against edited and composited photos. Pure AI-generated images have uniform compression (they were only saved once), so this signal alone won't catch them — but the other four will.
Every physical camera sensor produces a unique noise signature — random fluctuations from the electronics and photon physics. This noise is consistent across a real photo and has characteristics that vary by camera. SafPix isolates and analyzes this noise layer using established signal-processing techniques.
Many AI-generated images either have no real sensor noise (suspiciously "clean") or carry artificial noise that's inconsistent across the frame. It's one of our better signals for spotting pure AI generation — but we're honest that the newest generators are getting better at faking convincing grain, so SafPix weighs it alongside everything else rather than treating it as the last word.
Real photos have fine micro-texture everywhere — skin pores, hair strands, fabric weave, background detail. AI generators often produce unnaturally smooth regions, especially on faces and skin, because they optimize for visual plausibility rather than pixel-level realism.
SafPix maps the texture density across the image and flags regions that are too smooth to come from a real camera. You can see this yourself in the Texture overlay — real photos have detail everywhere, AI images have conspicuous "painted" zones.
The fifth pass runs a battery of statistical checks that look for patterns unique to AI generation — including color characteristics, structural symmetry, and pixel-level uniformity that real cameras don't produce. Each check alone is a weak signal, but when several fire together, the compound signature is strong evidence.
This category also includes an open-source AI-image detector that runs entirely in your browser — a small neural model trained on millions of real and AI-generated images. We're honest about its ceiling: no public AI detector is right every time, and the best ones sit around 80% on freshly released generators. So SafPix treats the model as one weighted vote, not a verdict — and when its different views of an image disagree, it abstains and says "unclear" rather than guessing. We keep adding checks as generators evolve.
How the signals combine
Each signal scores 0–100. SafPix combines them using a weighting system that adapts based on what evidence is available — giving more weight to the signals that matter most for each image. When multiple signals flag the same image, compound penalties strengthen the verdict. When an image carries intact Content Credentials, that cryptographic proof takes priority over the statistical signals — and when the evidence genuinely conflicts, SafPix would rather tell you it's unclear than force a confident-but-wrong answer.
A single weak signal could be noise. Three signals all saying "something's off" is a pattern. SafPix trusts the pattern — and abstains when there isn't one.
Most single-model detectors give you one percentage. When it's wrong, you can't tell why. SafPix runs 18 independent signals across these categories — plus the in-browser AI detector — with plain-English explanations and visual overlays you can toggle and inspect yourself.
The limits — honestly
We'd rather tell you what we can't catch than sell you a false promise.
- Well-prepared fakes. If someone generates a synthetic image and specifically optimizes it to pass forensic analysis — adding realistic noise, fixing metadata, cropping to non-standard dimensions — some signals may miss it. Always use the visual overlays and your own judgment to cross-reference.
- Deepfake video frames. SafPix analyzes single images. A deepfake video screenshot might score OK if the individual frame is clean.
- Subtle retouching. A single removed blemish won't be caught. A face swap usually will.
- Brand-new generators. Each major AI model shifts the statistical fingerprint. We continuously retune, but the newest models may initially evade detection.
Have a photo you're not sure about?
Drop it in. You'll have an answer in seconds — a clear sentence telling you what we found and what it means. No signup, no upload to servers, no guesswork.
Check a photo now →