FotoForensics: How to Use Error Level Analysis (ELA) to Spot an Edited Photo

0

๐Ÿ”Ž Reverse Image Search

Image search techniques made simple: upload a photo or paste a URL and search 6 engines at once. 100% free, nothing stored.

๐Ÿ”’ Your image is never stored on our servers longer than 24 hours and is never shared.

Direct answer: FotoForensics error level analysis (ELA) re-saves a JPEG at a known quality and shows where the copy differs from the original. Areas that were edited and re-saved usually differ more than the rest, so they stand out on the ELA map. It is a useful clue, not proof: bright areas can also be ordinary edges, texture, or phone processing.

An edited photo rarely announces itself. A face can be swapped, a sign reworded, or a small object removed, and the result can look perfectly natural at normal size. That is why a forensic check is useful, not as a verdict on its own, but as one more piece of evidence alongside the picture’s history and its metadata. This guide is the tool-depth companion to our visual-signs guide: it does not re-teach how to spot warped backgrounds or mismatched shadows. For those checks, read how to tell if a photo is photoshopped first, then come back here for the pixel-level method.

In the sections below you will learn what error level analysis actually measures, how to run a FotoForensics ELA tutorial step by step, how to read ELA results without fooling yourself, what the JPEG% figure means, and where the method breaks down. The centrepiece is a worked example I ran on 9 November 2026: the same mountain photograph, unedited and then edited with a cloned patch of sky, run through the tool twice and compared in a table. No screenshots are used in this article; every reading is described in words, so you can repeat the test yourself and compare what you see.

What Is Error Level Analysis, and How Does FotoForensics Do It?

Error level analysis, usually shortened to ELA, is a way of looking at JPEG compression rather than looking at the picture itself. A JPEG does not store every pixel exactly. It stores an approximation, and it throws away more detail in some places than in others. When a photo is saved as a JPEG, each 8 by 8 block of pixels is compressed, and the amount of error left behind depends on the quality setting and on how busy that block is. A flat sky compresses cleanly. A detailed rock face does not.

ELA exploits that behaviour. It takes the JPEG, re-saves it at a known quality, typically around 90 to 95 per cent, and then compares the re-saved copy with the original, pixel by pixel. Where the two agree closely, the error is low and the ELA map is dark. Where they disagree, the error is high and the map is bright. A part of the photo that was edited and then saved again has usually been through a different compression history from the rest, so it should show a different error level. That is the whole idea, and it is also the reason ELA is so easy to misread: anything that changes compression behaviour can change the error level, whether or not anyone edited the photo.

FotoForensics is the best-known free website for running this check in a browser, and it is the tool used for every test in this guide. Its ELA tutorial states that similar edges should have similar brightness in the ELA result, and that similar surfaces and similar textures should have similar colouring. FotoForensics also warns that resaving a JPEG removes high frequencies, and that scaling a picture smaller can boost high-contrast edges. Those two sentences are the honest core of ELA: compare like with like, and expect edges to be bright. If you remember nothing else from this guide, remember that.

One more point matters before you start. ELA only works properly on JPEG files, and it works best on the original file, at full resolution. A screenshot of a photo, a PNG converted from a JPEG, or a copy downloaded from a messaging app has already been re-compressed, resized, or both. The compression history you are trying to read has been overwritten. Always ask for the original file when the answer matters, and work on a copy so you never re-save the evidence itself.

How Do You Run a FotoForensics ELA Photo Check?

Running an ela photo check takes about two minutes, and it needs no account and no software. Use the largest, clearest JPEG you can get, and do the steps in order so you do not destroy the file you are testing.

  1. Keep the original safe. Make a copy of the file and analyse the copy. Do not open the suspect photo in an editor, rotate it, or save it again before the check, because every re-save adds a new compression layer.
  2. Open FotoForensics. The front page offers a file upload and an image URL box. Uploading the file is the more reliable route; the URL option is useful when the photo is already online and you do not have the file.
  3. Upload the JPEG or paste its URL. Wait for the analysis page to load. The tool first shows the source image, with a menu of analyses on the left: Digest, ELA, Games, Hidden Pixels, ICC+, JPEG %, Metadata, Strings, and Source.
  4. Choose ELA. The ELA map appears in place of the source image. Study it at full size, not as a thumbnail. Use the zoom and colour-adjustment controls if the map is hard to read, and compare the ELA view with the source view side by side in your mind.
  5. Check JPEG % next. Switch the analysis menu to JPEG % and note the estimated last-saved quality. Write the number down; you will need it when you interpret the ELA map, because a very low or very high figure changes what “normal” looks like.
  6. Check Metadata. Switch to Metadata and note the file size, dimensions, encoding, and any camera, software, or editing-history fields. Note what is missing too: a social-media copy will usually have had its camera data stripped.
  7. Record what you saw before you decide. Write down, in plain words, which areas were bright, which were dark, and whether the bright areas matched a shape, an object, or an edge in the original. A written note stops you from remembering the map as more suspicious than it was.

If the upload fails, check the file first. The tool works with JPEGs; a WebP, HEIC, or PNG file may need converting, and converting it will change its compression history, so treat any result from a converted file with extra caution. A file larger than the tool’s limit, or a URL that blocks outside requests, will also fail, in that case, download the file yourself and upload it directly.

How Do You Read ELA Results Honestly?

Learning how to read ela results is less about spotting brightness and more about spotting difference. A bright patch on its own means very little. What matters is whether a patch is brighter or darker than similar material nearby, whether its shape follows an object or an edge in the photo, and whether the same pattern appears in the other checks you run. Work through the three comparisons below every time.

First, compare same-texture regions. Find two areas of similar material, two patches of sky, two areas of grass, two parts of the same wall, and compare their error levels. If one is clearly brighter than the other, and there is no reason in the photo for the difference, that is worth noting. If both are equally bright, the brightness is probably a property of that texture, not an edit. A mountain ridge will always be brighter than a flat sky, because it contains far more detail for the compressor to work on.

Second, compare edges with edges. High-contrast edges are bright in ELA by design, because a sharp change from light to dark is hard to compress cleanly. The outline of a person against a bright background, the edge of a roof against the sky, and text on a sign will all glow. That does not mean they were pasted in. The question is whether the edge is brighter than other edges of the same strength in the same photo. FotoForensics re-saves the image and measures the difference, so an edge that is unusually bright compared with its neighbours is the signal, not brightness on its own.

Third, look for shapes, not blobs. A genuine local edit often leaves a shape with a purpose: a rectangle where a sign was covered, a repeated patch where a texture was cloned, a face or object whose outline is sharper than everything around it. Random bright speckles across a detailed area are usually just texture. A straight edge, a repeated pattern, or a region whose boundary matches an object in the original photo is far more interesting. If you cannot describe the suspicious area in words, where it is, what shape it has, and what it lines up with, you probably do not have a finding yet.

Finally, change the reading conditions before you conclude. Zoom in, adjust the ELA brightness, and look again. A real difference survives a change of view; a trick of the thumbnail does not. If a bright area disappears when you zoom in, it was probably a display artefact. If it stays, and it is confined to one area with a clear shape, keep it on your list and check it against the metadata and the reverse search.

What Does JPEG% Actually Tell You?

The JPEG % view is the part of the tool most people skip, and it is often the most useful. It estimates the quality at which the file was last saved as a JPEG, by comparing the image with the quantisation table that is stored inside the file. In a jpeg error level analysis online check, that number tells you what “normal” should look like for this particular file.

A high figure, such as 95 to 99 per cent, usually means the file was saved once at high quality, or that it was exported from an editor at a high setting. A mid-range figure, such as 80 to 90 per cent, is common for photos that have been through a website, a stock library, or a phone export. A low figure, below 70 per cent, suggests heavy compression, often after several saves or after a messaging app squeezed the file. None of those numbers proves editing. They describe the file’s history, and they set your expectations: a heavily compressed file will have a noisy, bright ELA map everywhere, so a single bright patch is weaker evidence in that file than in a clean, high-quality one.

JPEG % is also useful when it disagrees with the story. A photo that is supposed to be an untouched camera original, but which reports a last-saved quality of 65 per cent and lists editing software in its metadata, deserves a second look. A photo that reports 92 per cent and shows no software field is not proved genuine, the fields can be edited, but the two facts together are consistent with a lightly handled file. For a full explanation of the fields involved, including EXIF, IPTC, and XMP, see our guide to what image metadata is and how to read it.

The Metadata view completes the trio. It lists the file size, pixel dimensions, colour profile, and the internal JPEG structure, such as whether the file uses progressive encoding. More importantly, it shows whether any camera or software information survived. In our worked example below, the edited copy carried no camera data and no software name at all, only the basic JPEG and colour information. That tells you the test photo had already been through a stock library’s export pipeline before it reached the tool, which is exactly why the ELA reading had to be judged on its own terms, not against an ideal camera original.

What Did Our Worked Example Actually Show?

This is the ranking asset for this guide: a real test, run twice, on the same photograph, with one deliberate edit between the two runs. The test photo was a 1,200 by 800 pixel JPEG of a mountain sunrise above a sea of cloud, taken from Unsplash (photo 1506905925346-21bda4d32df4). I chose it because it contains both a flat, smooth sky and a highly detailed mountain ridge, the two textures that make ELA easiest and hardest to read in the same frame. The edit was a clone: I copied a patch of plain sky from the centre of the same photo and placed it, 200 by 140 pixels, over the small pink cloud at the top right, at position x=1000, y=70. The patch itself had been saved at a lower quality (60 per cent) before it was placed, and the final edited copy was saved at the same 85 per cent quality setting as the original. Both files were then submitted to FotoForensics by URL on 9 November 2026. No screenshots of the ELA maps are shown here; the readings below are the actual readings, described in words.

ReadingOriginal (unedited)Edited copy (cloned sky patch)
JPEG % (last-saved quality)89%, JPEG Standard89%, JPEG Standard
File in the Metadata view1,200 by 800 JPEG, as submitted by URL1,200 by 800 JPEG, JFIF 1.02, sRGB, progressive DCT, YCbCr 4:2:0; no camera or software fields
ELA: skyDark and low in error across the whole sky, with only faint specklesDark across the sky, as before, the patch interior was not brighter than the surrounding sky
ELA: mountain ridge and snow edgesBright white along the ridges, snow edges, and the rocky foregroundBright white in exactly the same places; unchanged by the edit
ELA: cloudsMid-grey and speckled, brighter than the sky but duller than the ridgeMid-grey and speckled, as before, except where the patch covered one cloud
ELA: the edited areaNo rectangle and no straight edge in the top-right sky; the pink cloud was visible in the sourceA thin rectangular outline appeared in the top-right sky, tracing the left and bottom edges of the patch; the pink cloud beneath it was gone
What the test provesThe bright ridge is normal texture, not editingELA flagged the patch boundary, not a bright fill, a subtle signal that is easy to miss

Three lessons follow from that table, and they are the reason the test is worth more than a generic warning. First, JPEG % alone would not have caught this edit: both files reported exactly 89 per cent, because the final save used the same quality setting. Second, the much-feared “bright blob” did not appear. The inside of the patch stayed dark, just like the sky around it, because a smooth sky compresses cleanly wherever it came from. What gave the edit away was the outline, a straight, rectangular edge in a part of the picture where nature draws no straight lines. Third, the rest of the map did not change at all, which is itself a check: if an ELA map changes everywhere after a small edit, something else, such as a re-save at a different quality, has happened to the whole file.

Be honest about the limits of this test, too. The patch I used was deliberately visible as a rectangle in the source photo itself, because its sky tone did not quite match the sky it covered. A careful editor who blends a patch properly would leave a fainter boundary, and a very small, well-blended clone in a busy texture might leave no ELA signal you could trust. This test shows what a clear ELA signal looks like when it is there; it does not show that ELA always finds edits. That distinction is the point of the guide.

What Are the False Positives That Make ELA Lie?

False positives are the reason ELA has a bad reputation among people who have only tried it once. They are not rare exceptions; they are normal, and they follow patterns you can learn. The table below covers the three the brief for this guide required, high-contrast edges, phone HDR and portrait processing, and heavy recompression, plus two more that catch beginners out.

False positiveWhy it lights upHow to avoid being fooled
High-contrast edgesSharp light-to-dark changes are hard to compress, so edges, text, and outlines are bright in almost every ELA mapCompare the edge with other edges of the same strength in the same photo, not with flat areas
Phone HDR processingPhones merge several exposures, sharpen, and add local contrast, which changes compression in parts of the frameExpect patchy brightness in HDR photos, especially around skies and windows; look for shapes, not patches
Portrait mode and background blurThe phone blurs the background and sharpens the subject, so the two areas have different compression historiesDo not compare the blurred background with the sharp subject; compare background with background
Heavy recompressionEach re-save adds error, and messaging apps and social platforms re-save at low qualityGet the original file; if you cannot, treat a bright, noisy map as uninformative rather than suspicious
Resizing and screenshotsDownscaling boosts high-contrast edges and discards detail, while a screenshot is a new photo of a screenWork at full resolution, and never judge a screenshot’s ELA map as if it were the original photo

There is a useful rule behind all five rows. If the bright area is explained by something you can see in the photo, an edge, a texture change, a blurred background, a sign, then the brightness is probably explained too. An edit, by contrast, is usually a difference with no visible reason in the source photo, or a difference whose shape matches an object that should not be there. Write the visible reason down next to each bright area. If you cannot, that area stays on your list; if you can, cross it off.

When Does ELA Fail, and What Should You Check Instead?

ELA fails in three common situations, and knowing them in advance saves you from both false alarms and false reassurance. First, it fails on files that have been re-saved so many times that the whole map is noise. Second, it fails on images that were never JPEG photographs in the first place, PNG graphics, screenshots, and AI-generated images have no JPEG compression history to expose. Third, it fails on careful, well-blended edits in busy textures, where the changed area compresses almost exactly like its surroundings. A clean ELA map therefore never proves a photo is genuine. It only proves that this one test found nothing.

When ELA cannot answer the question, use the checks that answer different questions. A reverse image search finds earlier copies of the same photo, which is often the fastest way to settle an edit: if an earlier, unedited version exists, the comparison is decisive. Our guide to how to tell if a photo is real puts ELA alongside six other verification methods, including source and date checks, so you can weigh them together. And for newer photos, check provenance before pixels: how to check Content Credentials (C2PA) on a photo explains the signed record that some cameras and editors now attach. The Content Credentials specification is a project hosted by the Coalition for Content Provenance and Authenticity (C2PA), according to Content Credentials, and its official site is contentcredentials.org. A photo with a valid provenance record that matches its story is far stronger evidence than any ELA map, and a photo whose record says it was generated or heavily edited should be treated accordingly.

For news, marketplace, and dating-profile photos, the practical order is: reverse search first, metadata second, ELA third. That order is the reverse of what most people try, and it is deliberate. The reverse search finds the original when one exists, the metadata tells you what the file says about itself, and ELA then tells you whether one area of that file behaves differently from the rest. No single step carries the verdict; the three together usually do.

Which ELA Tool Should You Use in 2026?

FotoForensics is not the only way to run error level analysis, and it is worth knowing what the alternatives do differently. The comparison below is based on the tools the leading forensic guides describe, including the FixThePhoto guide to checking edited images and the JPG.now forensic analysis guide published in August 2026.

Tool or methodWhat it is good atWhat to watch for
FotoForensicsFree ELA, JPEG %, and metadata views in one place, with no account neededELA maps need careful reading; bright edges are normal, as the worked example shows
Forensically (browser tool)Free, with ELA, clone detection, and noise analysis in one workspaceMore controls mean more ways to misread a result; use one analysis at a time
FotoForensics alternatives on JPG.nowGuides that pair ELA with quantisation tables, recompression generations, and AI-detection signalsA forensic guide is not a verdict; its checklist still needs a human to weigh the results
Editors such as PhotoshopUseful for visual checks, clone-stamp repeats, liquify warping, and layer edges are covered in our companion guideVisual signs alone cannot prove or disprove an edit; use them with a reverse search

JPG.now’s forensic guide, published in August 2026, lists treating ELA brightness as proof of editing as its first common forensic mistake. That warning is the best one-line summary of this whole guide: ELA is a screening tool, not a courtroom exhibit. In 2026 it is also a tool that has to sit alongside newer checks. AI-generated images, Content Credentials, and platform provenance labels all answer questions ELA cannot answer, and a photo can now be entirely synthetic while having a perfectly ordinary ELA map. Use ELA for what it is good at, spotting a local re-save in a genuine JPEG, and use the other checks for everything else.

Frequently Asked Questions

Is FotoForensics free to use?

Yes, the basic ELA, JPEG% and metadata checks are free to use in a browser, with no account and no software to install. You can upload a file or paste an image URL. There is a paid professional version with extra analyses, but the free checks are enough for most everyday verification.

Can ELA prove a photo was edited?

No. ELA shows where compression differs, not why it differs. A bright area can be a genuine edge, a texture, phone processing, or a re-save, as well as an edit. Treat it as a clue that needs a second check, such as the original file, metadata, or a reverse image search, before you make any claim.

Why is my whole ELA map bright?

A bright map usually means the file was heavily compressed, resized, or saved several times, so error is high everywhere. It can also happen with a screenshot, a messaging-app copy, or a very detailed photo. Get the original file at full size if you can, and compare only similar areas within the same map.

Does ELA work on AI-generated images?

Not in the way people hope. An AI image was never a JPEG photograph, so there is no original compression history for ELA to expose. Its map may look uniform or noisy, but neither result proves it is real or fake. Check for provenance records and reverse-search the image instead of relying on ELA alone.

FotoForensics is a good first pass, not a final answer. Run it on the original file, compare like with like, note the JPEG% and the metadata, and then check the photo’s history with a reverse search before you decide what to believe. If the ELA map, the metadata, and the history all point the same way, you have a case. If they disagree, trust the history, and do not share a photo you cannot verify.

Ali Butt is the lead writer and SEO specialist at Image Search Techniques, where he covers reverse image search, visual search, and online image verification. With over four years of experience in SEO and content, he writes practical, hands-on guides that help readers trace a photo to its source, identify people, products, plants and brands from a picture, protect their own images online, and spot catfish, fakes and AI-generated images across Google, Yandex, Bing, TinEye and more.

Leave A Reply