# What error level analysis shows, and what people think it shows

Source: https://passreal.com/detectors/fotoforensics-ela

Error level analysis re-saves a JPEG at a known quality and displays the difference between that and the original. Bright regions have compressed differently from their surroundings, which usually means they were edited or pasted. ELA does not detect AI generation, and reading a uniformly bright ELA image as proof of anything is the most common mistake made with the tool.

## What the tool actually computes

Error level analysis takes a JPEG, saves it again at a fixed quality, and
displays the per-pixel difference between the two versions. That is the whole
operation.

JPEG compression is lossy in a way that depends on what has already been thrown
away. A region compressed several times is already close to what the encoder
would produce, so re-saving changes it little and it appears dark. A region
introduced recently - pasted from elsewhere, painted in, rendered on top - has
not been through the same history, so re-saving changes it more and it appears
bright.

The output is therefore a map of compression history, not of authenticity. It
answers "has this part of the image been through the same processing as the
rest", which is a useful question when someone has spliced a helicopter into a
photograph, and a mostly irrelevant one when the entire frame came out of a
model in a single pass.

## Why generated images confuse people who read it

A generated image is internally consistent. Nothing was pasted, so nothing has a
different history, so there is no bright region standing out against a dark
background. The ELA view frequently looks calm.

The opposite failure is more common. Uploading a PNG, or a JPEG saved at maximum
quality, produces a frame that lights up everywhere, because the file has
effectively no compression history to compare against. People read that uniform
brightness as an alarm. It means the file has not been through a lossy encode,
which is true of every PNG ever made, including photographs exported as PNG.

## The tab that does the work

FotoForensics shows metadata alongside ELA, and for most uploads that tab
decides the matter first. It lists what the file declares about itself: make,
model, exposure, lens, software.

A file with no camera fields at all is unusual, because phones and bodies always
write them. A file carrying a software tag naming a generation tool has answered
the question outright. Neither finding requires interpreting a single pixel, and
both are faster to reach than any analysis.

## Reading an ELA view without fooling yourself

A few habits keep the tool useful rather than misleading.

- Compare against a control. Run a photograph you took yourself through the
  same view, at the same quality, before drawing conclusions about anything
  else. Without a baseline, every ELA image looks alarming.
- Ignore uniform brightness. It reports compression history, not manipulation,
  and a never-compressed file is uniformly bright by definition.
- Look for boundaries, not brightness. The finding is a region whose response
  differs sharply from its surroundings along an edge that does not correspond
  to anything in the picture.
- Expect high-contrast edges to glow. Text, logos and hard boundaries always
  respond strongly, in authentic photographs as much as anywhere else.
- Re-upload the same file twice. FotoForensics re-encodes on upload, so an image
  already analysed once has a different history the second time.

Practitioners treat the view as a pointer toward a region worth examining with
other methods. Read as a verdict on its own, it produces confident conclusions
about untouched photographs, which is how it acquired its reputation.

## Where processing fits

The pipeline's last stage writes the output as a JPEG with EXIF describing a
specific body, and with that make's measured quantization table rather than the
standard library one. The compression history becomes ordinary - one encode,
with tables consistent with the claimed origin - and the metadata tab reads like
a photograph's.

Upstream of that, the sensor stage restores a noise floor at intensity 0.018
and applies lateral chromatic aberration reaching 1.25 px at the corners,
and the frequency stage attenuates upsampling residue at 0.88 of
Nyquist. Those stages address statistical classifiers rather than ELA, which is
looking for something else entirely.

## Questions

### Does ELA detect AI-generated images?

Not directly. It surfaces inconsistent compression history within one file, which is a splicing signal. A generated image is internally consistent because the whole frame was produced at once, so it often looks unremarkable under ELA.

### Why does my PNG look completely white in ELA?

Because a PNG has never been JPEG-compressed. The first re-save differs everywhere, so the whole frame lights up. This says the file is not a JPEG, and nothing else. The same happens for a JPEG saved at very high quality.

### What are the other tabs on FotoForensics for?

The metadata tab is usually more informative than ELA. It lists EXIF, and for many uploaded files the absence of camera fields, or the presence of a software tag naming a generation tool, settles the question before any pixel analysis begins.

### Does processing change the ELA result?

Yes, though not as a goal in itself. The output is a JPEG encoded once with a camera quantization table, so it has a normal compression history and a normal ELA response, rather than the uniform brightness of a never-compressed file.

### Is ELA still used by professionals?

Rarely as evidence, often as triage. Its limitations are well documented in the forensics literature, and it is easy to misread. Practitioners treat a bright region as a reason to look closer, not as a finding.

Try it: https://passreal.com/app (three images free)