Image Diff

about this tool

Four ways to look at it

Side by side is the plain comparison. Slider puts one over the other with a handle to wipe between them, which is the fastest way to see a layout shift. Onion skin fades the second over the first, so a small movement shows as a doubling. Difference marks every changed pixel in red on a faded copy of the first, and that is the view worth saving.

What counts as different

Pixels are compared in YIQ rather than RGB, following the metric pixelmatch took from Yee's paper on perceptual image comparison. In that space brightness carries most of the weight and hue much less, which is how an eye reads a picture: a photograph nudged a shade darker looks changed, while the same size of shift in hue often does not.

The sensitivity setting is a fraction of the largest difference the metric can produce, so a lower number marks more. A transparent pixel is compared as it looks over white rather than as its stored colour, which stops a change in a fully transparent area registering as a change.

Anti-aliased edges are ignored by default. Text and curves are drawn with soft edges, and redrawing them a fraction of a pixel over changes thousands of pixels without changing anything anybody would notice. The check asks whether a pixel is an extreme among its neighbours in one image but not the other, which is what a moved edge looks like and what an actual change does not. Switch it off when you want every pixel.

Different sizes

Only the area the two share is compared, measured from the top left, and the panel says what that area is. Nothing is scaled to fit: scaling would resample every pixel and produce a difference everywhere, which tells you nothing.

Limits

There is no alignment: two screenshots taken at different scroll positions will differ everywhere, and the tool cannot tell you they are the same picture moved. Colour profiles are ignored, so the same image saved with different profiles may compare as changed. Both images are held in memory as raw pixels, four bytes each, so a pair of very large photographs will use a lot of it. Nothing is uploaded and nothing is stored between visits.

For two versions of a document rather than a picture, text diff compares them line by line. If the two images differ only in file size, image compressor shows what a given quality setting costs, and EXIF viewer reveals metadata differences a pixel comparison cannot see.

questions

What counts as a difference?

Each pair of pixels is compared in YIQ, where a change in brightness weighs more heavily than a change in hue, because that is how an eye reads it. The sensitivity setting is the fraction of the largest possible difference a pair has to exceed before it is marked, so a lower number catches more.

Why are some obvious changes not marked?

Probably the anti-aliasing option, which is on by default. Text and curves are drawn with soft edges, and a redraw at a slightly different position changes thousands of edge pixels without changing anything a person would notice. Switch it off to see every pixel that differs.

The two images are different sizes.

Only the area they share is compared, measured from the top left, and the panel says which area that is. Nothing is scaled: scaling to fit would change every pixel and make the comparison meaningless.

What are the four views for?

Side by side to read them both. The slider to wipe between them, which is the quickest way to spot a layout shift. Onion skin to fade one over the other, which shows a small movement as a doubling. And difference to mark every changed pixel in red on a faded copy, which is the one to save.

Are my images uploaded?

No. Both are decoded and compared in this page, and the difference image is drawn here too. Nothing leaves your device and nothing is stored between visits.