Is This Image Photoshopped?

How to Check if an Image is Photoshopped

Photoshop detection is fundamentally about finding the forensic inconsistencies that human editing introduces into the pixel data. It is a different problem from AI detection - and requires different techniques. This page focuses specifically on detecting manual editing: cloning, compositing, retouching, and manipulation by a human using tools like Photoshop, GIMP, or Affinity Photo.

Error Level Analysis (ELA) is the primary technique for detecting JPEG edits. Here is how it works: when you save a JPEG, the entire image is compressed at one quality level - every region of the image has been compressed the same number of times. When someone edits part of that image - pastes in a new element, removes a person, alters a background - and saves it again as JPEG, the edited region has now been compressed twice while the original regions have been compressed once. This difference in compression history creates measurable differences in how each region responds to further compression. ELA makes this visible as a brightness map: edited regions appear brighter because they retain more compressible error than the surrounding pixels that have already been fully compressed. FauxLens runs ELA automatically and overlays the result on your image so you can see exactly where the editing occurred.

Clone detection identifies duplicated pixel regions - areas copy-pasted from one part of the image to another. This is the most common technique for removing objects (copy a patch of background and paste it over the object you want to erase) and for adding objects (copy an element from another photo and composite it in). FauxLens identifies these duplicated patches and marks them in the forensic report.

Shadow physics consistency analysis checks whether shadows in the image obey the physical laws of optics. When elements are composited from different photographs, their shadows often come from different light sources - different angles, different distances, different softness. The tool checks shadow directions, ambient occlusion at object boundaries, and surface reflections for physical consistency. A professionally lit portrait composited into an outdoor background is one of the most detectable manipulations through shadow analysis.

Frequency domain analysis using Fourier transforms identifies the spectral signatures that modern Photoshop tools like Generative Fill and Content-Aware Fill introduce. These AI-assisted editing tools, now built into standard Photoshop, leave different artifacts than pure manual clone-stamp work. FauxLens detects both.

Photoshop vs. AI Generation - What is the Difference?

If you suspect an image has been manipulated, the first forensic question is: was a human editor involved, or was this image created entirely by AI? The answer determines which signals to look for and what the manipulation means.

Photoshopped images always start as real photographs. A human being used editing tools to alter specific regions - adding or removing elements, blending photos together, retouching faces, altering text or numbers in the frame. The forensic signature of human editing is localized: the ELA map will show bright spots in the areas that were touched and normal levels everywhere else. The rest of the image has the noise distribution of a real camera capture. The metadata may be intact from the original capture. Clone detection may reveal specific copied patches.

AI-generated images have no camera origin at all. They were synthesized mathematically from a text description or a reference image, and every pixel was computed rather than captured. The forensic signature is global: GAN fingerprints and frequency domain anomalies appear throughout the entire image, not in localized patches. The image typically has no genuine EXIF metadata from a camera.

The practical significance is different too. A Photoshopped image tells you that a real event or scene was deliberately altered - someone wanted to change what a real photograph shows. An AI-generated image tells you there was no event, person, or scene at all - the entire thing is fictional.

FauxLens detects and reports both types separately, with different evidence signals for each. The forensic report will clearly indicate whether the signals found are consistent with manual editing, AI generation, or both - which can occur when real photographs are edited using AI-powered tools like Photoshop Generative Fill.

For most people asking "is this Photoshopped?", they want to know whether a real photo was altered. The ELA and clone detection layers of FauxLens answer exactly this question, independent of the AI detection layers.

Limitations of Photoshop Detection

Honest disclosure of what forensic tools can and cannot catch is as important as describing their capabilities. Photoshop detection has genuine technical limitations that are worth understanding before interpreting a result.

JPEG quality level is the dominant variable in ELA effectiveness. When a photo is edited and re-saved at very high JPEG quality (Q95 or above), the difference in compression history between edited and original regions is smaller, making the ELA brightness map less distinct. Professional retouchers who know about ELA specifically sometimes save at high quality to suppress the signal. Conversely, images saved at lower quality levels (Q70 or below) produce very strong ELA contrast that makes edits easy to locate. Most consumer-shared photos fall between Q80 and Q90, where ELA works well.

PNG format has no JPEG compression and therefore no ELA signal at all. PNG is lossless, meaning every save produces identical pixel data. ELA is fundamentally inapplicable to PNG. For PNG images, FauxLens relies entirely on clone detection, shadow physics analysis, and frequency domain forensics. Detection accuracy for manipulated PNGs is lower than for manipulated JPEGs.

Uniform global adjustments do not trigger ELA. If someone applies color grading, curves adjustments, exposure changes, or contrast modifications to the entire image uniformly, ELA will not flag it - because the entire image has been processed uniformly. Only localized edits that affect one region differently from the rest produce detectable ELA signals. This means heavily color-graded images, heavily filtered photos, and images processed through Instagram or similar filters may show reduced ELA effectiveness even if they were additionally edited.

Very high-resolution compositing by experienced professionals is the hardest case. A professional compositor working from source TIFFs - never touching JPEG - who blends at the pixel level with custom feathering and matches lighting carefully before a single JPEG export may produce an image where ELA signals are minimal and clone detection finds nothing. The signals that remain most reliable against expert editing are shadow physics (lighting geometry is extremely hard to fake across composited elements from different sources) and skin texture uniformity (professional-grade face retouching produces texture uniformity that is statistically anomalous compared to real skin).

High-Profile Cases When Photoshopped Images Had Real Consequences

The stakes of undetected image manipulation are not hypothetical. Several high-profile cases in recent years illustrate what forensic detection can and should catch.

The Kate Middleton family photo incident (March 2024) became an international news story when the Princess of Wales released an official family photograph for Mother's Day that was later admitted to have been edited. Multiple news agencies - AP, Reuters, and Getty - withdrew the photo from distribution after identifying that elements of the image had been digitally altered. The specific manipulations included inconsistencies in the clothing and positioning of family members that were inconsistent with a single-shot capture. ELA and geometric consistency analysis would have flagged the compositing at the clothing boundaries. The incident sparked significant media coverage about image authentication standards in royal communications.

The World Press Photo controversy (2024) involved a Brazilian photographer whose winning image was disqualified after forensic examination revealed that elements had been composited from different frames - a practice that violates World Press Photo's standards for documentary photography. The disqualification was based on RAW file analysis and pixel-level comparison. Shadow geometry inconsistencies and compression artifacts at the composite boundary were the specific signals that initiated the investigation.

Sports photography manipulation has been documented in multiple instances where crowd size, equipment branding, or celebration poses were altered to benefit sponsors or create more dramatic imagery. These manipulations typically involve cloning background elements or replacing surfaces - detectable through clone detection and ELA.

Political campaign photo alteration is documented across multiple jurisdictions: crowd size manipulations to make rallies appear larger, opponent photos altered to appear more unflattering, and setting composites that place politicians in more or less favorable visual contexts. ELA on the background regions and shadow geometry analysis at the subject-background boundary are the primary detection signals for these manipulations.

The forensic signals that would have revealed each of these cases - ELA at composite boundaries, shadow direction inconsistencies, clone detection in background regions - are exactly what FauxLens runs automatically.

What Professional Photo Editors Do Differently - and How to Detect Both

Professional photo editors are more skillful at suppressing forensic signals than amateurs, but they are not immune to detection. Understanding what professionals do differently - and what signals remain reliable despite expert editing - is essential for accurate interpretation of forensic results.

Amateur editing leaves strong signals. An amateur using Photoshop on a JPEG will typically: open the JPEG, make their edit, save as JPEG. This produces strong ELA artifacts because the edited region has now been through two JPEG compression cycles. The amateur may also use obvious clone stamp patches that are detectable as exact pixel duplicates. Clone detection and ELA both fire clearly.

Professional editing is more forensically opaque. A professional compositor typically works in a non-destructive workflow: source files are RAW or TIFF (lossless), all editing happens in layers, and the final JPEG is exported in a single step at the end. This single-pass JPEG export means the entire image has been compressed exactly once, uniformly, which dramatically reduces ELA signals. The edited regions have never been through multiple JPEG cycles.

Additionally, professional retouchers blend edits at the pixel level with custom feathering and noise matching. Rather than pasting a hard-edged clone stamp, they use soft brushes and frequency separation techniques that match the surrounding texture. Clone detection is significantly less effective against frequency-separated retouching than against hard-edged stamp work.

However, several signals remain reliable even against professional editing. Shadow geometry is extremely difficult to fake when compositing subjects from different light sources. The angle, softness, and direction of shadows are governed by physics - a subject photographed outdoors at noon composited into a studio portrait will have shadow inconsistencies at the body boundaries that spatial light analysis can detect. Optical physics violations - reflections on surfaces that do not match the surrounding environment, ambient occlusion patterns at object boundaries that violate the apparent light field - similarly resist professional suppression.

Skin texture uniformity is another reliable signal. High-end beauty retouching produces statistically anomalous skin texture - areas that are too smooth, with too-low variance across large facial regions. Real skin has microscopic texture variation that is statistically measurable. Heavily retouched skin shows abnormally low texture variance in the retouched regions. FauxLens checks for this statistical anomaly.

The practical takeaway: for amateur and intermediate Photoshop editing, ELA and clone detection are highly reliable. For professional-grade compositing and retouching, shadow physics analysis and texture statistics are the signals that remain most informative.