Image diff tool — compare two images pixel by pixel

Use cases

Why compare two images pixel by pixel?

The eye is good at scenes and bad at pixels. Two versions of the same image can look identical on screen while differing in thousands of individual values — and a small, deliberate edit can hide in a busy photo. An image diff tool turns those invisible changes into something you can see and measure.

Three situations come up most often:

• Verifying metadata cleaning. Run a photo through the MetadataRemover metadata cleaner, then diff the original against the cleaned copy. The pixels should be identical — or, if you enabled the fingerprint reset, only a faint, evenly scattered glow from the invisible perturbation that changes the image fingerprint. Anything else would mean the cleaning touched your image content.
• Spotting edits and retouching. Compare a photo before and after someone worked on it. Object removal, skin retouching, cloned regions and doctored screenshots light up on the difference heatmap even when the result looks completely natural at a glance.
• QA for designers and developers. Check what a new export preset actually changed, measure the cost of a lower JPEG quality, or confirm a re-rendered asset still matches the approved reference. The percentage of changed pixels and the PSNR give you a number, not a guess.

Under the hood

How the difference metrics work

The tool draws both images on a local canvas and reads their pixels back. If the dimensions differ, only the overlapping region — measured from the top-left corner — is compared, and the note above the views tells you the exact size of that region. The visual modes are unaffected: each image is still shown at its own scale.

Four numbers describe the pixel difference. Pixels changed is the share of pixels where at least one RGB channel differs by more than the sensitivity threshold you set with the slider. Mean absolute difference averages the per-channel absolute error across the compared region — a low value means the images are close even where they differ. PSNR (peak signal-to-noise ratio) compresses the mean squared error into decibels: higher is closer, 40 dB and above is visually indistinguishable for most content, and identical pixels give an infinite PSNR, which the stats panel reports explicitly instead of a number.

One caveat worth knowing: lossy recompression changes almost every pixel a little. Saving the same photo as JPEG twice at different qualities will light up the heatmap everywhere at low thresholds. Raise the threshold to ignore that compression noise and surface only real edits.