Find what changed between two versions of an image. Add A, then B, and the differing pixels are counted and drawn. Four views cover the two questions that follow: where is the change, and what is it.
Highlight and heat
Highlight draws A in pale grey and paints each changed pixel red. It is a yes-or-no map: a pixel is over the tolerance or it is not. The readout above gives the count, the share of all pixels, and below the picture the bounding rectangle of everything that changed, as a size and a top-left position you can crop to.
Heat colours every pixel by how far it moved, from black for none through purple and orange to pale yellow for a complete change. The scale is compressed so small shifts are visible. A recompressed JPEG shows as a dim haze across the whole frame with brighter block edges, and a swapped icon shows as one bright patch. Heat ignores the tolerance, because it shows magnitude instead of a decision.
The download is the view you are looking at, Highlight or Heat, at the full size of the comparison.
Swipe and onion
Swipe shows A to the left of a divider and B to the right. Drag on the picture, or use the slider, to move the divider across a region and watch it change.
Onion lays B over A at an opacity you set. At 50% anything that moved shows as a double image, and running the slider back and forth makes a small shift flicker.
How a difference is measured
Each pair of pixels is converted from RGB to YIQ, a colour space that separates brightness from two colour axes, and the squared differences are weighted 0.5053, 0.299 and 0.1957. Brightness counts most because the eye is most sensitive to it. Partly transparent pixels are compared as they would appear over white.
Tolerance is a fraction of the largest possible distance, applied to the squared value. 10% corresponds to about a 1% squared distance, a difference that is hard to see with both images side by side.
The anti-aliasing test looks at the eight neighbours of a differing pixel. If the pixel lies between a darker and a brighter neighbour, and one of those neighbours sits in a flat area of three or more identical pixels in both images, it is an edge pixel that two renderers smoothed differently. Those are painted yellow and not counted.
Example: two screenshots of one page
Take a screenshot before and after a CSS change, at the same window size. Add both. With the defaults, a changed button colour shows as a solid red rectangle and the readout gives its bounds, for example 184 × 44 at 912, 630. If the whole lower half is red, the page reflowed: something above changed height. Switch to Swipe and drag across the top of the red area to find the element that grew.
If text edges across the page light up yellow and nothing is red, the two captures rendered fonts differently and nothing else changed.
Shifted captures
A capture that is one pixel off flags every edge in the image. Open Nudge B and move B right or down by a pixel at a time until the changed count collapses. The offset that gives the lowest count is the shift between the two captures, and what remains red is the real difference.
Limits
The comparison is positional. A photo that was rotated, cropped from a different corner or taken from a slightly different angle differs everywhere, and no setting will line it up. Use the Image Hash Comparator to ask whether two such images show the same subject.
Pairs larger than 24 megapixels are compared at reduced size, and the readout shows the scale. Counts then refer to the reduced pixels.
Lossy formats differ from their source even when nothing visible changed. Comparing a PNG with a JPEG of it at tolerance 0 flags most of the image; raise the tolerance until the haze clears and what is left is structure.