Guide · 8 min read · Updated September 2026

Signal three: editing signatures, and how a real photo shows its scars

Most manipulated media is not generated from nothing. It is a real photograph with something removed, something added, or something changed — a person deleted from a group, a figure moved, a number altered, a face swapped onto a real body. That work leaves a different class of evidence from generation, because the tell is not that the image looks synthetic; it is that one region of an authentic photograph has a different history from the rest of it.

By Ira Peoples

The principle: local history mismatch

A photograph accumulates a consistent history. One sensor recorded it with one noise character. One lens applied one pattern of sharpness and vignetting. One compression pass divided it into blocks and quantised them identically everywhere. Whatever happened to it, happened to all of it.

An edit breaks that uniformity in one place. The pasted or painted region carries either the history of a different image or no history at all, and every editing-signature test is a way of asking the same question: does this patch belong to the same photograph as its surroundings?

Compression history

JPEG encoding works on an 8×8 grid, quantising each block by a fixed table. Save, edit a region, and save again, and the untouched areas have now been compressed twice against the same grid while the edited region has effectively been compressed once. That difference is measurable as an uneven error level across the frame, and it is the basis of error level analysis.

It is a genuine signal and a badly misused one. Different materials legitimately compress differently — a flat sky and a detailed tree will never show the same error level — so a bright region in an ELA map means 'compressed differently', not 'faked'. The finding worth anything is a region whose compression character has a shape: a rectangle, or an outline that follows an object rather than a texture.

Noise discontinuity

Every sensor produces noise, and its amount and character vary predictably with brightness and with the camera's ISO setting. Across one photograph the noise is statistically uniform for a given brightness level.

A region taken from another photograph brings that photograph's noise, which will differ if the two were shot at different ISO or on different cameras. A painted or generated region often brings noise that is too clean or synthetically even. Areas that are suspiciously noiseless amid natural grain, or a patch where grain size visibly changes, are among the most reliable editing signals — and unlike compression tests they are hard to fake deliberately, because matching noise convincingly requires effort most manipulators do not make.

Resampling and the softened boundary

Pasted elements almost always need resizing or rotating, and both require resampling — recomputing pixel values by interpolation. Interpolated pixels are mathematically related to their neighbours in a way original sensor pixels are not, which introduces detectable periodic correlation.

By eye, the visible consequence is the edge. A real object boundary is as sharp as the lens allows; a composited one is often a hair softer, because the manipulator feathered it to hide the seam. Trace suspicious outlines at high magnification and look for a boundary that is smoother than every other boundary in the frame, or a faint halo where the edited region's contrast was adjusted.

Cloning and content-aware fill

Removing an object usually means filling its place with material copied from elsewhere in the same image, either by hand with a clone stamp or automatically. The filler is real photograph, so it has the right noise and the right compression — but it is duplicated.

Look for repetition that nature would not produce: two identical clusters of leaves, a paving pattern that repeats exactly, a stretch of wall or water with a copied irregularity. Look also for broken continuity where the fill met a structure it could not reason about — a railing, a shoreline, or a line of text that does not resume correctly on the other side of the patch.

Colour and lighting mismatch

Colour grading is applied to a whole image, so hue, saturation and contrast should be coherent across it. An inserted element frequently carries the white balance of its source: a face slightly warmer than the body it sits on, an object with black levels that do not match the scene's.

Check the darkest and brightest points of a suspicious region against the darkest and brightest points elsewhere. If the region's blacks are lifted or crushed relative to the rest, it was graded separately — which is to say, separately sourced.

The limits, stated plainly

These tests read a file's history, so anything that gives the file one new history defeats them. Screenshotting an edited image, re-saving it heavily, or passing it through a social platform re-compresses everything uniformly and erases local differences. A skilled manipulator can also add matching noise and re-encode the whole frame deliberately.

And the biggest caveat is intent: editing signatures detect editing, not deception. Cropping, colour grading, retouching a blemish, straightening a horizon and removing a bin from a holiday photo all leave the same traces as a malicious change. The question the evidence answers is 'has this region been worked on?' — what that means is a judgement about the claim being made, which no tool can make for you.

The short version

Editing tests ask one question: does this region share the rest of the photo's history?
Noise discontinuity is the most reliable family; error level analysis is real but widely over-read.
A boundary softer than every other boundary in the frame suggests a feathered composite.
Repeated texture and broken structure mark cloning or content-aware fill.
Re-saving and resharing erase all of it — and detecting an edit is not detecting a lie.