Forensic Metadata Analysis
Inspects dimensional signatures, aspect-ratio integrity, format fingerprints and bit-per-pixel density to surface the inconsistencies that editing pipelines and generators leave behind.
Live100% in-browser · nothing uploaded
Know if an image is real in seconds.
Fake Images Detector runs a four-layer forensic sweep across metadata, sensor noise, compression behaviour and error levels — exposing AI-generated, deepfaked and digitally edited photos before you trust them. Free, unlimited, and it never sees your file.
- No sign-up required
- Zero-upload privacy
- Unlimited scans
ELA MappingRegion-level anomalies
ANOMALY · 0.87
NOISE · SYNTHETIC
Zero UploadProcessed on device
Built for the people who can't afford to be fooled
- Newsrooms
- Trust & Safety
- Legal & eDiscovery
- OSINT Analysts
- Academic Research
- Insurance Claims
The Detector
Drop an image. Get a forensic read-out.
Four analysis layers run simultaneously on your device. You get an authenticity score, a per-layer breakdown, an error level map that shows where an image was touched, and a colour-channel fingerprint.
AI Image DetectorDetect manipulated, deepfaked or AI-generated images
Your image never leaves this deviceDrop your image here, or click to browse
You can also paste from your clipboard with Ctrl / ⌘ + V
JPGPNGWEBPMAX 10MB
Analyzed Image
Authenticity Analysis
Authenticity Score
Error Level AnalysisELA
Low error · untouchedModerate · processedHigh error · likely edited
Color AnalysisRGB
RedGreenBlueGaps or spikes can indicate re-grading
These results are heuristic forensic indicators, not a legal determination. Read the score alongside the ELA map and the per-layer breakdown, and corroborate high-stakes findings with source verification.See our disclaimer.
Multi-Layered Detection
Four forensic layers, one clear answer
No single test catches every fake. Our engine runs four independent analyses in parallel and cross-references them, so a manipulation that slips past one layer is caught by another.
Noise Pattern Recognition
Real camera sensors produce chaotic, channel-independent grain. Diffusion models and retouching produce statistically smooth noise — we measure the difference across every colour channel.
Compression Artifact Intelligence
Re-encoding leaves quantisation traces. By profiling how an image responds to controlled re-compression we detect edit generations invisible to the naked eye.
Precision Error Level Mapping
The ELA visualiser renders a heat map of compression error across the frame, highlighting spliced objects, cloned regions and pasted faces as bright anomalies against a uniform background.
How It Works
Three steps, about two seconds
No account, no queue, no credit card. Drop a file and read the result.
- 1
Upload your image
Drag and drop, browse, or paste straight from your clipboard. JPG, PNG and WEBP up to 10 MB. The file is read into memory by your own browser — nothing is transmitted.
- 2
Instant multi-layer analysis
Metadata, noise, compression and error levels are computed in parallel on the HTML5 canvas. You can watch each layer complete in real time as the progress track advances.
- 3
Read the full report
You get a headline verdict, an authenticity score, a per-layer breakdown with plain-English explanations, an ELA heat map and a colour histogram — plus a downloadable text report.
Why Choose Us
Forensics you can inspect, not a black box
Most checkers hand you a percentage and expect you to trust it. We show the evidence: which layer flagged what, and where on the image it happened.
Evidence, not just a verdict
Every score is broken down into the four measurements that produced it, each with the raw figure behind it — so you can form your own judgement instead of deferring to ours.
Beyond basic edits
Detects GAN and diffusion output, neural rendering, face swaps and deepfakes, splicing and copy-move forgeries, plus subtle work like colour grading, exposure lifts and filter stacks.
Genuine zero-upload privacy
Analysis is pure client-side JavaScript. Sensitive evidence, unpublished photography and confidential case material never cross the network — which is why legal and newsroom users can actually use it.
Unlimited, and free for good
No credits, no throttling, no watermarked exports. Because the compute happens on your machine, there's no server bill to pass on to you.
Detection coverage
AI generation (GAN / diffusion)96%
Deepfake & face swap93%
Splicing & copy-move91%
Retouch & colour grading88%
Re-compression generations94%
| Category | Detection rate |
|---|---|
| AI generation (GAN / diffusion) | 96% |
| Deepfake & face swap | 93% |
| Splicing & copy-move | 91% |
| Retouch & colour grading | 88% |
| Re-compression generations | 94% |
Indicative detection rates from internal benchmark sets of original, edited and generated images. Real-world performance varies with image quality and how many times a file has been re-saved.
Trusted Worldwide
Used by researchers, journalists and investigators
5M+
Images Analyzed
Millions of scans run by journalists, moderators, insurers and investigators across more than 150 countries — every one of them processed locally.
4 Layers
Cross-Referenced Signals
Metadata, noise, compression and error level analysis are weighted together, so a single ambiguous reading never decides the outcome on its own.
24/7
Unlimited Access
No usage caps, no rate limits, no subscription tier. Free for personal and professional use, permanently and without registration.
Detection Coverage
What we catch
The engine pairs classical image forensics with modern generative-artefact detection to cover the full manipulation spectrum — from a crude paste-in to a photorealistic diffusion render.
- AI-generated imagery — GANs, diffusion models and neural rendering
- Professional editing traces at pixel level (Photoshop, GIMP, Affinity)
- Object insertion, removal and transformation
- Facial engineering, face swaps and deepfake artefacts
- Subtle enhancement — colour grading, relighting and filter stacks
- Multi-generation re-saves and screenshot laundering
The ELA map is what sold me. Being able to point at the exact region that fails and explain why to an editor — in a tool that never uploads the file — is worth more than another opaque confidence number.
We screen hundreds of claim photos a week. This is the first free checker our compliance team cleared, precisely because nothing crosses the network.
I teach a digital forensics module with it. Students can see the noise and compression layers disagree with each other, which is exactly the lesson.
How accurate is the fake image detector?
The detector combines four independent forensic layers — metadata integrity, noise pattern analysis, compression profiling and error level analysis — into a single authenticity score. Heavily edited and fully generated images are flagged reliably.
Lightly processed photographs sit in a grey zone by design: rather than forcing a false binary, the tool returns an inconclusive band and shows you which layer disagreed. Always read the score together with the ELA map.
Which manipulation techniques can it detect?
Signatures left by AI generators such as Midjourney, DALL·E and Stable Diffusion; professional editing software including Photoshop and GIMP; technical manipulations such as splicing and copy-move forgery; facial modification and deepfakes; and lighter processing like filters, colour and lighting adjustments.
Detection heuristics are updated as new generation techniques and their artefacts become documented.
Are my images uploaded to a server?
No. Every calculation runs inside your browser using the HTML5 Canvas API. There is no upload endpoint, no storage bucket and no transmission of image data — you can verify this yourself in your browser's network inspector.
When you close or reload the page, the image is gone from memory.
How well does it identify AI-generated content?
AI-generated images carry characteristic traits: unnaturally uniform noise across colour channels, an atypical response to re-compression and large stretches of near-identical neighbouring pixels. Those traits depress the noise and pixel-consistency layers, which pulls the composite score down.
Upscaled, filtered or screenshot-laundered generations are harder, because the laundering destroys the very artefacts we measure — always prefer the original file.
What are the ideal file specifications?
JPG, PNG and WEBP up to 10 MB. For the most reliable reading, use images between 800×600 and 4000×3000 pixels — that range gives the algorithms enough detail without drowning them in noise.
Smaller and larger files still analyse fine, but screenshots and social-media re-saves have already been re-compressed, which weakens every forensic signal.
Is it really free, and is there an API?
The web tool is completely free and unlimited — no account, no throttling, no watermarks. Since the analysis runs on your hardware, there is no cost for us to recover.
For batch or programmatic verification, get in touch and tell us about your volume and use case.
Can I use the results as legal evidence?
No. The output is a heuristic indicator intended to support human judgement, not a certified forensic finding. It carries no chain of custody and is not notarised.
For proceedings where authenticity is contested, engage a qualified forensic examiner who can produce an admissible report.
Which browsers are supported?
Any modern evergreen browser — Chrome, Edge, Firefox, Safari and their mobile equivalents. The tool needs HTML5 Canvas, FileReader and IntersectionObserver, all of which have been standard for years.
Because everything runs locally, analysis speed depends on your own device rather than on our servers.
Ready to see the truth behind an image?
Professional-grade image verification, free, unlimited and private by construction. No registration — just drop a file.