Error Level Analysis (ELA) Explained Simply

Error Level Analysis, or ELA, is one of the best-known image forensics techniques — and also one of the most misunderstood. It pops up in online debates the moment a viral photo is questioned, brandished as irrefutable proof of tampering. The reality is more nuanced. ELA is a useful tool, but it only works if you understand what it actually measures and, above all, what it does not. This guide demystifies the method, teaches you to read an ELA map without falling into the usual traps, and explains why it is only worth one piece of a much larger puzzle.

What is Error Level Analysis?

ELA is a forensic analysis technique that exploits a fundamental property of the JPEG format: "lossy" compression. When a JPEG image is saved, the algorithm discards part of the information to shrink the file. This loss is not uniform: it depends on the content of each small block of pixels.

The core idea of ELA is simple. If you recompress a JPEG image at a known quality level and then compare the result to the original, the areas that have already been compressed several times react differently from those compressed only once. In theory, a part of the image that was added, pasted in, or retouched — and therefore saved at a different "compression age" than the rest — stands out with a distinct error level.

Why JPEG is at the heart of the method

JPEG splits the image into 8×8 pixel blocks and applies a mathematical transform to each (the discrete cosine transform). The more times an area has been compressed, the closer it gets to a "stable state" where further recompression changes little. Conversely, a freshly introduced area is still "far" from that state: it changes more during recompression.

It is precisely this difference in reactivity that ELA tries to make visible. The JPEG format is therefore essential to the principle: on a PNG or an uncompressed raw file, ELA simply makes no sense.

A bit of history and context

ELA became popular in the late 2000s, driven by consumer tools that made it accessible in a single click. It had its heyday during controversies over press, contest, or social media photos, where amateur observers used it to "prove" tampering. This virality had a perverse effect: the method became both overused and misused. Many ELA-based "revelations" actually rested on a misreading of the maps, mistaking sharp edges for manipulation. Understanding the method therefore also means understanding why it produced so many hasty conclusions.

Compression, quantization, and "error level"

To go one step further without diving into equations: JPEG compression relies on a "quantization" step that rounds the frequency coefficients of each block. This rounding is the source of information loss. When you recompress an image, the blocks already heavily rounded move little (they are already "locked" onto the quantization grid), while content introduced with a different rounding history readjusts more. The "error level" measured by ELA is therefore nothing but the magnitude of this local readjustment. This nuance explains why ELA depends so heavily on the image's original JPEG quality and on the quality chosen for recompression.

How an ELA analysis actually works

An ELA analysis comes down to a few steps that can be summarized as follows:

1. Take the image to analyze (a JPEG).
2. Re-save it at a fixed, known quality level (often around 90–95%).
3. Compute, pixel by pixel, the difference between the original and this recompressed version.
4. Amplify that difference to make it visible: this is the "ELA map".

The result is an image where the brightness of each point represents the magnitude of the compression error at that spot. Dark areas changed little; bright areas changed a lot.

Reading an ELA map without getting it wrong

Here is the point most people miss: an ELA map is never read in the absolute, but by relative consistency. You are not looking for a "bright" or "dark" area in a vacuum. You are looking for an area that behaves differently from what its nature should imply.

A few reading cues:

• Sharp edges, contours, and rich textures always stand out more than smooth surfaces. That is normal — not evidence of tampering.
• A plain sky, smooth skin, or a wall naturally appear dark in ELA.
• The suspicious signal is an inconsistency: an object whose error level sharply differs from similar elements elsewhere in the image, or a contour that "glows" while everything else has been flattened by successive recompressions.

What ELA can detect

Used correctly, ELA can highlight several types of manipulation:

• Splicing: an element imported from another image often keeps a different compression history.
• Cloning / stamping: an area duplicated to mask an object may show an error signature inconsistent with its surroundings.
• Localized retouching: added text, a removed detail, an embedded logo.
• Partial multiple recompressions: when only part of the image was re-saved after editing.

ELA is therefore primarily a detector of local modifications on JPEG photographs. That is where it is most relevant, and its history is tightly linked to the analysis of news and contest photos.

What ELA CANNOT detect

This is where most interpretation errors concentrate. ELA has major blind spots you must know before drawing any conclusion.

• Fully AI-generated images: an image produced end to end by a generative model is uniform from a compression standpoint. There is no area "added" at a different compression age — everything was created at once. ELA is therefore largely blind to pure AI images.
• Global modifications: a change in brightness, contrast, or a filter applied to the whole image creates no local inconsistency.
• Heavily recompressed images: after several passes through social media, the image reaches a stable state where any trace of manipulation is drowned out. ELA then becomes unusable.
• PNGs, screenshots, and non-JPEG files: without original JPEG compression, the method has no working basis.

The false-positive trap

ELA generates many false positives, and that is its main weakness. A high-contrast area, a sharp edge, a signature, legitimately embedded text, or simply an object more detailed than its surroundings: all of these can "glow" without any fraudulent manipulation. Conversely, a skillful edit on an already heavily compressed image may leave no visible ELA trace. An ELA map is therefore never proof on its own.

Summary table: strengths and limits of ELA

A concrete example: reading an ELA map step by step

Let's take a typical case to anchor the theory. Imagine a cityscape photo onto which an advertising billboard has been added.

1. Global observation: on the ELA map, the sky and smooth façades appear dark (little error), while the building edges, windows, and foliage glow (lots of detail, hence lots of error). All of this is normal.
2. Spotting the anomaly: the billboard, however, shows an abnormally sharp contour and a uniform error level that contrasts with neighboring objects of the same type. That is a first signal.
3. Internal comparison: you compare this billboard to another billboard actually present in the original scene. If both should have a comparable signature but do not, the doubt grows.
4. Cross-checking: you look at whether the billboard's edge coincides with a "break" in the image's texture, a sign of splicing.

Note that none of this is proof. A billboard can legitimately have a different compression history without being a fake. ELA points to an area to examine; it is other methods that confirm or refute.

Classic reading pitfalls

• Confusing sharp edge = tampering. False: edges always glow.
• Believing a dark area = authentic area. False: a smooth surface is dark by nature.
• Reading a map without internal reference, in the absolute.
• Working on a file already recompressed by a platform.

Where ELA fits in a multi-layered analysis

The key lesson is that a serious analysis never relies on ELA alone. It is one layer among others, to be cross-checked systematically. Even before looking at an ELA map, you must examine the file's context: its EXIF metadata and what it reveals about an image tells you about the device, the date, and any passage through editing software.

To distinguish an authentic photograph from a synthetic image — where ELA is blind — you need other methods: analysis of the sensor noise (PRNU) that acts as a device fingerprint, the statistical signatures specific to generators, and AI vision detectors. The full approach is detailed in our guide on how to detect an AI-generated image by cross-checking every signal.

This is exactly the logic TruthLens applies: pixel-by-pixel ELA is only one brick of a pipeline that combines EXIF, C2PA, AI vision, generator signatures, and PRNU. The verdict never comes from a single test, but from the convergence of independent layers. You can submit an image to this full analysis in seconds and get a reasoned report.

Why a certified report changes everything

For evidentiary use — press, insurance, litigation — producing an ELA map is not enough. You need a document that holds up, dated and tamper-evident. A certified report with a SHA-256 fingerprint and timestamp guarantees the analysis was not altered after the fact. This is the topic of our guide on how to certify the authenticity of an image or video.

Best practices for using ELA

If you run ELA yourself, keep these principles in mind:

• Work on the original file, never a screenshot or a copy re-downloaded from social media.
• Never read an ELA map in isolation: compare areas of the same nature with one another.
• Distrust your intuition: edges and textures always glow — that is not a sign of fraud.
• Cross-check with metadata and at least one other forensic method.
• Document your reasoning: a defensible conclusion explains why an inconsistency is suspicious, not merely that it "glows".

Conclusion: a useful tool, not a magic wand

Error Level Analysis remains a wonderful teaching tool for understanding how JPEG compression betrays certain manipulations. But its reputation as a universal "Photoshop detector" is greatly overstated. It is blind to AI images, prone to false positives, and unusable on overly degraded files. Its real value emerges when it is integrated into a multi-layered analysis like the one TruthLens performs, corroborated by other signals, and recorded in a certified report. It is this methodological rigor — not an isolated test — that lets you seriously answer the question "was this image tampered with?".

FAQ

Does ELA prove an image was tampered with?

No. An ELA map highlights compression inconsistencies that may indicate manipulation — but may also correspond to perfectly legitimate sharp edges, textures, or text. ELA is a clue to corroborate, never standalone proof. A serious conclusion always cross-checks several methods.

Does ELA work on AI-generated images?

Very poorly. An image produced entirely by a generative model is homogeneous in terms of compression: no area was added at a different "age". ELA then has nothing to reveal. For AI images, you need other approaches: AI vision, generator signatures, watermarks, and noise analysis.

Why shouldn't you use a screenshot for ELA?

Because a screenshot re-saves the image into a new JPEG, erasing the original compression history. Any trace of earlier manipulation is drowned in this new uniform layer. ELA must always rely on the original, intact file.

What quality level should you use for ELA recompression?

There is no single ideal value, but a high quality (often 90–95%) is common because it amplifies differences enough without saturating them. What matters most is comparing areas of the same nature rather than trusting an absolute threshold, which has no universal meaning.