How it works
JPEG is a lossy format. Every time an image is saved, detail is thrown away in small blocks, and each save loses a little less than the one before, because the easy detail has already gone.
ELA exploits that. It saves the image again at a fixed quality and measures how much each part changed. An untouched photo that has been through the same saves everywhere changes by a roughly even amount across similar surfaces. A region that was pasted in from another image, or edited and saved a different number of times, changes by a different amount - and shows up brighter or darker in the ELA view.
The technique was popularised by researcher Neal Krawetz in the late 2000s and is now a standard first look in image forensics.
What bright areas actually mean
This is where most misreadings happen. Bright does not mean edited.
| You see | Usually means |
|---|---|
| Bright edges and outlines | Normal - sharp contrast always compresses with more error |
| Bright text and fine texture | Normal - high detail behaves the same way |
| Dark, flat sky or walls | Normal - smooth areas lose little |
| One object brighter than similar surfaces around it | Worth a closer look - possible different compression history |
| The whole image uniformly dark | Saved many times; ELA can no longer tell anything |
The useful comparison is always like with like: skin against skin, grass against grass. A pasted face that glows while the other faces in the photo do not is a lead. A bright edge is just an edge.
Where it fails
- Only JPEG. PNG and other lossless formats have no compression history to compare.
- Heavy recompression by social networks flattens everything, hiding edits and real differences alike.
- Careful editors can save an edited image in a way that evens the error levels out.
- AI-generated images are created whole, so there is often no local edit for ELA to find.
A clean ELA result therefore proves nothing, and a suspicious one is a reason to investigate, not a verdict.
SourceCheck: Detect Fake Photos & Deepfake
JPEG forensics with error level analysis, alongside C2PA verification, AI detection and metadata - on iPhone, iPad and Mac. · iPhone, iPad & Mac