How to Read ELA Heatmaps: A Plain-English Guide

If you have ever run an image through FakeRadar and wondered what the colourful heatmap actually means, this guide is for you.

What is Error Level Analysis?

When a JPEG image is saved, the encoder compresses each block of pixels (typically 8×8 pixels) using a process called discrete cosine transform (DCT). Each block is compressed to a slightly different degree depending on the detail in that area.

ELA exploits a simple idea: if you re-save an already-compressed JPEG, uniform parts of a genuine image will compress further at a predictable, consistent rate. If one region resists re-compression — or compresses at a markedly different rate — it suggests that region was saved at a different compression history than the rest.

In plain language: regions that stand out in an ELA heatmap have a different editing or saving history than the surrounding pixels.

What the Colours Mean

FakeRadar renders ELA results as a brightness heatmap overlaid on the original image:

• Bright / white regions — high error level, meaning that area was saved with a different compression history. Common in edited regions or pasted-in elements.
• Dark / black regions — low error level, meaning those pixels are well-settled after re-compression. Typical of flat areas like clear sky or solid backgrounds.
• Mid-tone regions — moderate error; common in fine textures (grass, hair, fabric).

A suspicious image often shows one or a few bright patches in an otherwise dark or consistent heatmap. That contrast is the key signal.

What a Genuine Image Looks Like

A genuine, unedited photograph will have a fairly uniform ELA pattern:

• Edges and textures (faces, foliage, text) appear slightly brighter.
• Flat areas (sky, walls) appear darker.
• No single region should stand dramatically above or below its surroundings unless it contains genuinely more detail.

Common Manipulation Signatures

Copy-paste or splicing

A region pasted from a different image will almost always have a different compression history. It may appear unusually bright (if pasted from a less-compressed source) or unusually dark (if pasted from a heavily compressed source).

AI-generated faces or objects

AI generators like DALL-E, Midjourney and Stable Diffusion produce images without any compression history — they start fresh. When saved as JPEG, the entire image shares one compression pass. This often results in an unusually uniform ELA pattern with no structural variation, which differs from the expected texture of a real photo.

Retouching and cloning

Skin smoothing, object removal, or clone-stamp work can leave subtle ELA signatures — usually appearing either uniformly darker than expected (over-smoothed) or brighter at the boundaries of the retouched area.

What ELA Cannot Tell You

ELA is a forensic signal, not a verdict. There are important limitations:

1. Multiple saves degrade accuracy. If a genuine image has been saved and re-saved many times (common with social media screenshots), ELA can produce misleading bright areas that look like edits.
2. PNG and lossless formats show no ELA signal. ELA relies on lossy JPEG compression artefacts.
3. High-quality AI image generators can fool ELA. Some newer models produce outputs that pass ELA inspection.
4. JPEG quality settings matter. Very high-quality JPEGs compress so little that ELA contrast is minimal; very low-quality JPEGs create so much artefact noise that signals drown.

How FakeRadar Uses ELA

FakeRadar combines ELA with four other signals — Hive AI’s deepfake classifier, FFT frequency analysis, C2PA content credential verification, and EXIF metadata inspection — and presents a composite signal report. No single signal determines the result.

This is the right approach. ELA alone is not enough. But as one layer in a multi-signal analysis, it adds meaningful forensic value — particularly for detecting spliced or copy-pasted content.

Summary

ELA is best understood as one piece of evidence in a larger investigation — not a pass/fail test on its own. Use it alongside the other signals FakeRadar provides, and read the results with appropriate caution.

Want to run ELA on your own image? Upload it to FakeRadar — it’s free.