This tool computes (dis)similarity between two PNG images using (my approximation of) algorithms approximating human vision.

conversion assumes D65 white point and uses gamma from PNG file, defaulting to 2.2.

Comparison is done in L*a*b* color space using a multi-scale variant of the SSIM algorithm.

It started as a rewrite of Rabah Mehdi's C++ implementation:

Source and download

Usage

dssim file.png file-modified.png

Will output something like 0.2341. 0 means exactly the same image, >0 (unbounded) is amount of difference.

dssim -o totally-rad-visualisation.png file.png file-modified.png

Will also output map of differences. Transparent areas will have funky pattern added.

Example

Differences between original and a compressed version of an image. Different scales of analysis pick different distortions (in case of JPEG: noisy edges, blurred out edges, blockiness and quantized color).