AI Image Forensics

Upload an image to see its metadata (EXIF, XMP, PNG text), C2PA presence, frequency spectrum, noise residual and ELA on one screen. All analysis runs inside this browser; the image is never sent anywhere.

Image

Loading image…
Drag and drop or paste with Ctrl+V also works

Findings

This tool gives no AI verdict. It lists the evidence it found and how strong each piece is. Metadata can be stripped or forged, so evidence that is present means far more than evidence that is absent.

Metadata and file structure

FFT frequency spectrum

Input

Low frequencies are at the center, high frequencies at the edges. Real photos fade smoothly from the center outward. Upsampling layers in generative models can leave regular bright dots (red circles). JPEG files show grid dots spaced N/8 apart (grey circles) from 8×8 block compression; those are not a sign of generation.

Noise residual

The original minus a 3×3 median-filtered copy. Camera photos carry sensor noise spread evenly across the frame, with texture that changes with brightness. Generated images may have almost no noise, unnaturally different texture between regions, or repeating patterns.

ELA (error level analysis)

Re-saves the image as JPEG and amplifies the difference from the original. In a JPEG that was saved once, a region that stands out much brighter may have been pasted in later. ELA is better at finding local edits than at detecting AI generation, and PNG files light up evenly everywhere.

What this tool cannot do

For C2PA it only checks presence and reads embedded strings; it does not verify signatures. You can verify signatures at contentcredentials.org/verify. Invisible watermarks such as SynthID need the vendor's own detector, so this tool cannot check them. Images passed through social media or messengers, or captured as screenshots, lose their metadata and get recompressed, which weakens most of the evidence.