image

Error Level Analysis

A forensic technique that resaves an image at known quality to expose regions edited or generated differently from the rest.

Error Level Analysis (ELA) works by resaving an image at a known JPEG quality and comparing the result to the original. Areas that have been edited, spliced, or generated separately compress differently from the surrounding pixels, so they show up as brighter or differently-textured regions in the ELA map.

ELA was originally a photo-manipulation forensic tool — it's good at revealing where something was pasted into a photo. It also has uses against AI images: an inpainted region (a face swapped into a real photo, an object added) often has a different compression history than its surroundings, and ELA can surface that boundary.

ELA is not a standalone AI detector — it produces a visual map that a human interprets, and it's noisy. But it's part of the forensic toolkit journalists and investigators use, and it's worth understanding because inconsistent compression is a real giveaway.

SynthGuard reduces ELA signal in two ways. The JPEG double-compression layer gives the whole image a single, consistent compression history rather than a patchwork. For composited or inpainted images, boundary-aware smoothing blends edited regions into their surroundings so they don't stand out under ELA.