Image Verifier & Forensics — Authenticity Check

Forensic image analysis — verify file format vs declared extension, extract full EXIF / XMP / ICC / IPTC metadata, run Error Level Analysis (ELA), histogram, pixel-anomaly detection, structural artifact check, JPEG quantization fingerprinting, C2PA / AI-generator signature scan, and compression history. 100% client-side.

Tool Media & Files Updated Apr 26, 2026
How to Use
  1. Drag and drop an image, or click to select from your device.
  2. Every check runs locally in your browser — your image is never uploaded.
  3. Read the verdict at the top, then drill into each analysis tab for details.
  4. Open the ELA tab on a JPEG: bright patches mark areas that were compressed differently from the rest.
  5. Check the Metadata and AI signatures tabs for the camera, editing software and any C2PA or AI-generation signature.
Drop an image to verify
Drop an image here, paste from clipboard, or click to select
Supports JPEG, PNG, WebP, GIF, BMP, TIFF, HEIC, AVIF, raw (.cr2 / .nef / etc.) — anything decodable. Up to ~50 MB.
Try it with a sample

What each test means

Identity
Magic bytes vs extension
First 8-16 bytes of the file uniquely identify the format. Mismatch = file was renamed.
EXIF / XMP
Embedded metadata blocks
Camera, lens, capture time, GPS, editing software trail.
ELA
Re-save delta visualization
Bright spots = areas with different compression history (often pasted / edited).
Histogram
Color / luminance distribution
Detects clipping, tone-curve manipulation, cast issues.
Noise variance
Laplacian σ across regions
Real photos have uniform sensor noise. AI / composites often show noise discontinuities.
JPEG quality
Quantization-table fingerprint
Estimated saved quality + camera fingerprint when known QT matches.
C2PA
JUMBF manifest
Adobe / OpenAI / MS embed signed credentials describing AI use, edits, original camera.
SD signature
PNG tEXt parameters
Stable Diffusion / ComfyUI write generation parameters into a PNG text chunk.

Limitations & honest caveats

  • ELA only works on JPEG. PNG / WebP / lossless formats won't produce useful ELA output.
  • AI-detection is heuristic. Models change weekly; detection lags. A green "no signature" result does NOT prove an image is real — only that the operator didn't leave a trail.
  • PRNU matching requires a reference set from the same camera. This tool reports sensor-noise uniformity only — useful as a tampering hint, not as identity verification.
  • Deepfakes / face-swaps need ML models too heavy to ship as a web tool. Use a dedicated forensics service for that level of analysis.
  • The verdict is a guideline. Don't use it as the sole basis for a legal / journalistic claim. Reproduce findings in another tool before publishing.

About

The history of image forensics traces back to the 1990s when courts started accepting digital photos as evidence. Hany Farid at Dartmouth (now Berkeley) is generally credited with formalizing the field — his lab built the first ELA, copy-move, and re-sampling detection tools in the early 2000s. The C2PA standard (2021) is the industry response to AI-generated content: a cryptographically signed provenance manifest that travels with the file across crops and re-saves.

This tool is not a replacement for professional forensics software (FotoForensics, Amped, or law-enforcement-grade Cellebrite). It's a fast, private first-pass that runs entirely in your browser, surfaces the obvious red flags, and gives you ground truth on what the file actually is — extension, format, embedded data, basic statistics — before you make any claims about it.

Worked Example

A renamed file. You are sent cat.png. The Identity tab reads the first bytes: FF D8 FF, not 89 50 4E 47. It is a JPEG with the wrong extension, which usually means it was renamed or passed through software that does not check.

Reading ELA. The tool re-saves the image as a JPEG at quality 0.95 and subtracts the result from the original. If a pixel’s red, green and blue differ by 3, 1 and 2 levels, the largest difference is 3, and the ELA image shows 3 × 6 = 18 out of 255, a dark grey. A pasted region that was compressed differently might differ by 20 levels and show as 120, clearly brighter than its surroundings.

The common mistake: reading a bright edge as a fake. Sharp edges, text and saturated colours always produce larger ELA values than smooth areas, even in an untouched photo. What matters is whether similar textures look different from each other: a bright rectangle on a smooth wall is suspicious; a bright outline around every edge is normal.

Show Work

Format by magic bytesFF D8 FF at offset 0 = JPEG, not PNG
Channel differences|R−R′| = 3, |G−G′| = 1, |B−B′| = 2 = largest is 3
ELA pixelmin(255, 3 × 6) = 18
Edited patchmin(255, 20 × 6) = 120
Ratio between them120 ÷ 18 = about 6.7 times brighter

Formulas

Error Level Analysis
ELA = min(255, 6 × max(|ΔR|, |ΔG|, |ΔB|))
Δ is the original minus the copy re-saved at JPEG quality 0.95
Noise estimate
σ of the Laplacian response per region
Compared across regions for consistency
Magic bytes
JPEG FF D8 FF, PNG 89 50 4E 47, GIF 47 49 46 38
The file’s real type, whatever its name says

How Image Forensics Works

JPEG compression works on 8 × 8 blocks and throws away fine detail each time an image is saved. A region that has been saved once responds differently to another save than a region that has been saved several times, which is the idea behind error level analysis. It only means anything on a lossy format such as JPEG.

The other checks look for traces rather than pixels. Cameras write EXIF metadata, editors add their own software names, content-credential standards such as C2PA embed signed records of edits, and some AI image tools write their prompts into PNG text chunks. Any of these can be stripped, so a clean result proves little, while a positive finding is solid evidence.

Frequently Asked Questions

What does "claimed format vs actual format" mean?

Every file format has a unique signature in its first bytes (magic bytes). PNG starts with 89 50 4E 47, JPEG with FF D8 FF, etc. If a file is renamed (cat.jpg → cat.png) or generated by software that lies about format, the magic bytes won't match the claimed extension. This tool reads the actual bytes and reports any mismatch.

What is Error Level Analysis (ELA)?

ELA re-saves a JPEG at a known quality level then compares pixel-by-pixel against the original. Areas with consistent error (same compression history) appear uniform; areas that were recently pasted, edited, or had different compression show as noticeably brighter regions. Originally invented to spot photo manipulation. Effective on JPEG only — PNG / lossless formats yield no signal.

Can it detect AI-generated images?

Sort of. The tool looks for: (1) C2PA / Content Credentials manifests embedded by Adobe Firefly, OpenAI, Microsoft, etc. (2) XMP / EXIF software fields naming AI tools (3) PNG tEXt / iTXt chunks containing Stable Diffusion-style prompts and parameters (4) Statistical signatures common to current generative models. Detection is improving but never definitive — a careful operator can strip every signature.

What is PRNU?

Photo-Response Non-Uniformity — the unique noise fingerprint a camera sensor leaves on every image it captures. PRNU matching needs reference images from the same camera, so a single-image tool can't verify "this is from camera X" without that database. What this tool reports is sensor-noise uniformity: whether the noise pattern is consistent (likely a real photo) or has discontinuities (likely composite or AI).

Are my images uploaded?

No. Everything runs in JavaScript in your browser. The image never leaves your device. The page works offline once cached.

How do I use the Image Verifier & Forensics — Authenticity Check?

Simply type your numbers and read the result, which refreshes the instant you change something. There is nothing to submit and nothing to wait for.

Is it free? Does it work without internet?

Yes to both. It is free with no sign-up, and once the page has loaded it keeps working even with no internet.

Where does my data go?

Nowhere — every calculation runs on your own device. Nothing you enter is uploaded, logged, or stored.

Common Use Cases

News / journalism fact-check

Check whether a viral photo has been edited. ELA spotlights pasted regions; metadata reveals if it was processed.

Catfish / dating-app vetting

See if a profile photo was downloaded from elsewhere (compression history) or AI-generated (signature scan).

Insurance / fraud claim review

Identify staged or manipulated damage photos via ELA and metadata cross-checks.

Stock photo authenticity

Confirm a "real photo" claim by checking camera EXIF and absence of AI signatures.

Forensic chain-of-custody

Document an image's declared format, hash, dimensions, EXIF, and compression history for legal records.

Last updated: