ImageComparisonSettings Class

The distance is the largest value of the Butteraugli map of libjxl (ButteraugliDistance.MaxDistance), a psychovisual model that gives a distance for every pixel: unlike the SSIM and PSNR means, a difference confined to a small area weighs fully. Images that are not fully opaque are composited over black and over white, and every pixel keeps the larger distance. It is the slowest check. Images must have the same dimensions.

The distance is the Hamming distance between the hashes, plus the difference between the mean luminances of the images on the same scale (a solid black and a solid white image are 64 apart; each unit is about 4 luminance levels out of 255), capped at 64. A bit is set only when the left thumbnail cell is brighter than the right one by more than half a luminance level, so invisible noise in flat areas does not flip bits. Images that are not fully opaque are compared composited over a black and over a white background, and the larger distance is used. Images must have the same dimensions.

The distance is computed like the one used by ImageComparisonSettings.DHashThreshold. A bit is set only when its DCT coefficient exceeds the median coefficient by more than one luminance level, so invisible noise in flat areas does not flip bits.

The ratio is computed from the mean squared error of the R, G and B channels premultiplied by alpha, the value scikit-image's peak_signal_noise_ratio(expected, actual, data_range=255) computes for opaque images. The alpha channel gives its own ratio, and the lower of the two is used. Like SSIM, it is a mean: a localized difference lowers it less than the same difference spread over the image. Images must have the same dimensions.

The score is the mean SSIM over every 7×7 window of the images (uniform weights, sample covariance, K1 = 0.01, K2 = 0.03), the value scikit-image's structural_similarity computes with its default parameters. The R, G and B channels are premultiplied by the alpha channel and averaged, and the result is the lower of that score and the score of the alpha channel. Because the score is a mean, a localized difference lowers it in proportion to the area it covers: a difference confined to 1% of the image lowers it by about 0.01 at most. Images must have the same dimensions.

The score is the one of the ssimulacra2 tool of libjxl, a perceptual metric that compares the images in the XYB color space at 6 scales and penalizes new artifacts (ringing, banding, blockiness) and lost details (blurring). Images that are not fully opaque are composited over a dark and a bright gray, and the lower score is used. The metric needs images of at least 8×8 pixels: smaller images pass this check only when their pixels are identical (fully transparent pixels are equal whatever color they hide). Images must have the same dimensions.