Free, private AI image check—no upload
Drop a supported image below to inspect noise, JPEG-grid, texture, entropy, channel, frequency, EXIF and dimension signals. The analysis happens on your device and explains conflicting evidence instead of hiding it behind one label.
Check an imageright here.
No login. No upload. Drop any image and run all eight forensic signals in your browser — the score and full breakdown are free.
Drop an image here or click to browse
JPEG, PNG, or WebP only
All 8 signals run locally in your browser. No image data is uploaded.
One label is notan answer.
Commercial detectors hand you a single number from a black-box model. When it's wrong, you have nothing to argue with. SynthGuard shows you every signal that contributed to the verdict.
DCT block analysis, edge entropy, colour-channel correlation.
FFT periodicity, noise distribution, JPEG quantisation tables.
EXIF consistency, software signature, output geometry.
Six pillars,one verdict.
Every probe is grounded in physical sensor properties or compression mechanics — not a single trained model that breaks the moment a new generator ships.
Noise Consistency
Real CMOS sensors scatter photon shot noise unevenly across the frame. Diffusion models smooth it. We measure the variance and call out the give-away patches.
8×8 Compression Grid
JPEGs from real cameras leave a deterministic 8×8 quantization grid. AI exports often skip or double-quantize it. We probe the block-edge regularity and read it like a barcode.
Local Detail Variance
Patch-level texture variance tells you where the model gave up. We sample dozens of windows across the frame and flag the suspiciously uniform ones.
Color Distribution Entropy
Generators tend to collapse hue distributions into narrow gaussians. We compute Shannon entropy on the histogram and surface the anomalies.
Channel Correlation
On real sensors the red and green noise channels correlate predictably. AI output decorrelates them. That R/G covariance is a single number you cannot fake by accident.
EXIF & Output Geometry
Camera, lens, GPS, timestamps, software signature — and tell-tale generator dimensions like 1024², 768², 832×1216. Eight signals fused into one verdict.
Eight probes,one fused score.
Each probe returns a status (safe / warning / risk) and a weighted point contribution. The final 0–100 verdict is a weighted fusion of these contributions, calibrated against thousands of real and generated samples.
- 01Noise consistency map
- 028×8 JPEG block boundary probe
- 03Local detail variance sampler
- 04Color distribution entropy
- 05R/G channel correlation
- 06FFT high-frequency energy
- 07EXIF reliability audit
- 08Output geometry classifier
What eachprobe measures.
A complete reference of the eight forensic signals — what they look at and what the two extremes mean.
| Signal | What it measures | Reads as |
|---|---|---|
| Noise consistency | Sensor noise variance across the frame | Sensor / Generator |
| 8×8 JPEG grid | DCT block-boundary regularity | Real / Synthetic |
| Local detail variance | Patch-level texture entropy | Natural / Smoothed |
| Color entropy | Histogram Shannon entropy | Wide / Collapsed |
| Channel correlation | R/G noise cross-covariance | Coupled / Decorrelated |
| High-frequency energy | FFT roll-off at high bands | Crisp / Rolled-off |
| EXIF reliability | Camera / lens / GPS / software | Trusted / Suspect |
| Output dimensions | Known generator geometries | Camera-like / Generator |
Every scoreshows its work.
Click any indicator and you get the underlying measurement, its expected range for real photographs, and the weighted points it contributes — positive or negative — to the final verdict.
No "trust the model". Every probe is a deterministic computation you can re-derive from the pixels yourself.
Read the GAN-fingerprint explainer
Three steps,no friction.
Sign in once, drop any image, read the verdict. No credits, no queue, no upload.
Drop image
Drag any JPEG, PNG, or WebP image onto the canvas. The file stays on your device.
Local analysis
Eight forensic signals run inside your browser in 2–5 seconds. No upload, no queue.
Read the verdict
A 0–100 score plus per-signal breakdown. Click any indicator for the full reasoning.
Built for peoplewho need certainty.
Verify submitted photographs before publication.
Triage AI-generated uploads on user platforms.
Spot-check student-submitted imagery for AI use.
Flag synthetic product photos in listings.
Initial forensic pass before deeper analysis.
QA your own outputs before they hit the feed.
Your imagesnever leave your device.
We built the detector browser-only on purpose. No upload buckets to subpoena, no temp folders to forget, no ML training silently happening on your evidence.
Pixel data stays in browser memory. We literally cannot see it.
All eight probes run on Canvas + WebCodecs in your tab.
We have no dataset, because we have no images.
Nothing persists. Refresh and it's gone.
Related tools
Strip watermarks, inject PRNU, beat image detectors.
Raise burstiness and perplexity to pass GPTZero & Turnitin.
Per-frame humanization for AI-generated video.
Strip C2PA, SynthID and other AI-origin watermarks from any image.
Strip GPS, camera and timestamp data from any photo — free.
Questions,answered.
Is the AI Image Detector free?
Yes — it consumes zero humanize credits and has its own free weekly pool (10 scans/week on Free, 200 on Pro, unlimited on Studio). We offer it because the same forensic stack also powers our Photo Humanizer — knowing what you're up against makes the humanization more effective.
Are images uploaded anywhere?
No image pixels are ever uploaded — the full eight-signal pipeline runs in your browser using Canvas + WebCodecs. For signed-in users we log a run event with basic metadata (file name, size, score) so we can show your history and rate-limit abuse; your image data itself never leaves the device.
How accurate is it?
Treat the score as a forensic indicator, not a courtroom verdict. Clean diffusion output usually pushes the score into the elevated/AI-leaning range, but heavily edited, re-compressed, or EXIF-stripped images shift lower because real and synthetic signals overlap — which is exactly why we expose all eight indicators individually instead of a single number.
Which generators does it catch?
It can surface signals associated with diffusion or GAN output, including unusual noise, frequency, texture and metadata patterns. Those signals overlap with edited real photos, so the tool cannot reliably identify every generator or prove authorship.
Why eight signals instead of one model?
A multi-signal view makes it easier to inspect why a score changed and where evidence conflicts. It still needs human judgment and can be wrong when new generators or editing workflows change the signal distribution.
Can I use it on my phone?
Yes. The pipeline runs on iOS Safari, Chrome Android and any modern mobile browser. Per-image analysis takes roughly 4–8 seconds on phones versus 2–5 on desktop.
What file formats are supported?
JPEG, PNG, or WebP. Images are decoded natively in the browser — nothing is sent to a server. iPhone HEIC and AVIF aren't supported yet; convert to JPEG or PNG first.
How is this different from Hive or Sightengine?
Commercial detectors run server-side, take your image, and return a single black-box probability. SynthGuard runs locally, exposes every signal, and doesn't keep your file. Same forensic concepts, different threat model.
Stop guessing.Start measuring.
Open the detector in your browser. Drop one image. Read the verdict.