Image diff

What the tolerance does

A pixel counts as changed when any of its red, green or blue values differs by more than the tolerance. At zero you see every difference, including the invisible ones that JPEG compression introduces. Around 10 ignores compression noise and shows real edits. Raise it further to find only substantial changes.

What it is good for

• Checking whether a build changed a screenshot, which is the whole idea behind visual regression testing
• Seeing exactly which region of a page moved
• Confirming an export really is lossless
• Spotting edits in a photograph

The same thing in Pillow

from PIL import Image, ImageChops

a = Image.open("before.png").convert("RGB")
b = Image.open("after.png").convert("RGB")

diff = ImageChops.difference(a, b)
box = diff.getbbox()

if box is None:
    print("identical")
else:
    print("they differ inside", box)
    diff.save("diff.png")

getbbox() returning None means the two are pixel for pixel identical, which is the quickest test there is.

Counting the changed pixels

from PIL import Image, ImageChops

diff = ImageChops.difference(a, b).convert("L")
mask = diff.point(lambda v: 255 if v > 10 else 0)
changed = sum(mask.point(lambda v: v // 255).getdata())

print(changed, "of", a.width * a.height, "pixels differ")

A measure rather than a count

from PIL import ImageStat

stat = ImageStat.Stat(ImageChops.difference(a, b))
print(stat.mean)
print(stat.rms)

Root mean square across the whole image is the usual single number for "how different are these", and it is what most automated comparisons threshold on.