AI ImageDetector
Upload a local image and review 8 browser-based signals. Files never leave your browser, and only signed-in users can access this route.
AI Image Detector is sign-in only
Eight forensic signals run entirely in your browser — but you need a free account to start an analysis session. No credits used, ever.
3 free credits every Monday. No card.
100% browser-only · your files never leave your deviceFree account · no credit card required
Detector pillars
What the eight signals actually measure
Noise & Sensor Forensics
Real cameras leave a noisy fingerprint with channel correlations diffusion models can't reproduce cleanly. We measure noise level and channel correlation to flag synthetic textures.
8x8 JPEG Grid Analysis
JPEG-encoded camera files have block-boundary regularity at the 8x8 grid. We measure that ratio and downweight it automatically when the image was downscaled and the original grid no longer applies.
High-Frequency Energy
Diffusion outputs are smoother than reality at fine scales. An FFT-based pass measures high-frequency texture complexity to catch over-smoothed regions invisible to the eye.
Color Entropy & Variance
Generated images often collapse into narrower color distributions and over-uniform local patches. We score color entropy and patch-level variance against natural-photo baselines.
EXIF Reliability Tiers
Metadata is graded High / Medium / Low based on what's actually present — camera, lens, GPS, timestamps, software. Pixel signals are reweighted accordingly so missing EXIF can't dominate the verdict.
Local-first, weekly quota
No upload, no logs. The full eight-signal analysis runs in your browser. Free plan: 10 detector scans / week. Pro: 200 / week. Studio: unlimited under fair use.
How it runs
Three steps, all in your browser
Drop image
Drop a JPEG, PNG, or WebP image anywhere on the page. The file stays on your device.
Eight signals run
Pixel passes (noise, JPEG grid, FFT, entropy, channel correlation, patch variance) plus EXIF reliability. Cancel anytime.
Read the verdict
0–100 likelihood score plus a per-signal breakdown showing exactly which checks pushed the score up or down.
Detector signals
Synthetic vs. authentic — what each signal looks like
-
Noise consistencyOver-smoothedNatural range
-
8x8 JPEG gridWeak / missingCamera-typical
-
Patch varianceUniformNatural variation
-
Color entropyNarrow paletteBroad spread
-
R/G channel correlationUnusualSensor-like
-
EXIF reliabilityMissing / syntheticCapture metadata
FAQ
Detector FAQ
Is the detector accurate?
It's a forensic aid, not a verdict. The 0–100 score combines eight independent signals with weights that adapt to EXIF reliability. Use it together with context — no public detector is bulletproof against the latest generators.
Why is my real photo flagged as AI?
Heavy social-media re-compression, screenshots, and aggressive editing strip the very signals that prove authenticity (EXIF, original JPEG grid, sensor noise). Try the original file straight from your camera if you have it.
Why is an AI image scored as real?
Img2img, upscalers, photo-style LoRAs and post-processing through real EXIF can mask diffusion artefacts. We surface every signal individually so you can spot suspicious EXIF or unnaturally low channel correlation even when the total score is moderate.
Does this cost credits?
The detector has its own weekly quota separate from humanize credits. Free: 10 detector analyses / week. Pro: 200 / week. Studio: unlimited under fair use. The humanize pool is never touched by detector runs.
Does my image get uploaded?
Never. Decoding, EXIF parsing and all eight signal passes run in your browser. Your image bytes never leave the device — the only network call is a lightweight quota check so we can rate-limit abuse.
Which formats are supported?
JPEG, PNG, or WebP. iPhone HEIC and AVIF aren't supported yet — export or convert the image to JPEG or PNG first, then run the detector.