point the camera at an image
Trust what you’re looking at.
Green where nothing is found.
Red where generation is detected.
50 checks a month for a dollar. Cancel whenever you like.
Trust what you’re looking at.
Getting ready. Your photos are analysed and deleted straight after — never stored, never used to train anything.
GREEN — the detector thinks this was made by a camera.
RED — it thinks this was generated by an AI model.
NOT SURE — the signal is too weak to call either way.
This is a statistical guess, not proof. It is trained mostly on Stable Diffusion / SDXL-era images, so it is weakest on the newest generators, on heavily edited real photos, and on screenshots. Pointing a camera at a screen is the hardest case of all — re-photographing destroys the fine pixel texture the detector reads, so expect more NOT SURE there than when you open a file directly.
Treat green as "nothing suspicious found", not as "verified real". Absence of evidence isn't evidence of absence.
Where it exists, cryptographic provenance (C2PA / Content Credentials, signed by the camera at the moment of capture) is far more reliable than pixel forensics. It only works on original files, not on a camera pointed at a screen — so it complements this rather than replacing it.
When you check an image it is sent over an encrypted connection to our detection provider, analysed, and deleted immediately. It is never stored, never shown to anyone, and never used to train a model. We keep no copy and no account of what you check.
SMOGY AI-image detector — Swin, 224px, int8-quantised ONNX, 93 MB. Running via transformers.js on ….