BlazeDiff vs OpenCV

The blazediff Python wheel is ~3.3x faster than reading two files with OpenCV and running cv2.absdiff, and it answers the question a visual test asks: how many pixels differ, and where. OpenCV gives you a difference array, and the counting, thresholding and anti-aliasing handling are up to you.

cv2.absdiff subtracts two arrays per channel and returns the absolute values. That is a building block, not a test assertion: it has no perceptual threshold and no anti-aliasing detection, so one-pixel font rendering changes count as differences. blazediff.compare decodes both files, runs a YIQ color delta with optional anti-aliasing detection, and can write a diff image.

At a glance

Benchmarks

The OpenCV side reads both files as grayscale, runs cv2.absdiff, and counts the non-zero pixels. The BlazeDiff side reads both files in color and runs compare. Decode is included for both.

Switching

# Before
import cv2
import numpy as np
 
a = cv2.imread("baseline.png", cv2.IMREAD_GRAYSCALE)
b = cv2.imread("current.png", cv2.IMREAD_GRAYSCALE)
changed = np.count_nonzero(cv2.absdiff(a, b))
 
# After
from blazediff import compare
 
result = compare("baseline.png", "current.png", "diff.png", threshold=0.1, antialiasing=True)
if not result.match and result.reason == "pixel-diff":
    print(f"{result.diff_count} pixels differ")

compare returns reason == "layout-diff" when the two images differ in size, where cv2.absdiff raises an error.

When to keep OpenCV

• The diff is one step in a larger vision pipeline: alignment, contours, morphology or feature matching.
• You need formats that only OpenCV reads, such as TIFF, WebP or HDR.
• You need GPU processing.

Related

• Python API
• BlazeDiff vs pixelmatch, which covers the PyPI pixelmatch port
• Image difference analysis for regions instead of a count
• All benchmarks