BlazeDiff vs dssim

The blazediff-ssim crate computes MS-SSIM ~6.6x faster than dssim, and it is MIT-licensed, while dssim is AGPL-3.0 or commercial. The two answer related but different questions. dssim is a perceptual quality metric for tuning image codecs. BlazeDiff SSIM is a similarity score for visual tests.

dssim computes multi-scale SSIM in L*a*b* color space, scales in linear light, and reads color profiles. Its score is 1/SSIM − 1, so 0 means identical and the value has no upper limit. blazediff-ssim computes luma SSIM that follows the MATLAB reference implementations, and it scores from 0 to 1.

At a glance

	dssim	BlazeDiff (`blazediff-ssim`)
License	AGPL-3.0, or a commercial license	MIT
Delivery	Rust crate, CLI, C library	Rust crate, Node addon (`ssim-native`), JavaScript, Python
Input	PNG and JPEG in the CLI; RGBA in the library	RGBA in the crate; PNG, JPEG and QOI in `ssim-native`
Color	Lab* with chroma, color profiles	Luma; `perceptual-ssim` adds Lab* chroma
Score	DSSIM, 0 = identical, unbounded	SSIM, 1 = identical
Variants	One multi-scale metric	SSIM, MS-SSIM, Hitchhiker’s SSIM, perceptual SSIM
Threads	Multi-core by default	Single-threaded; you run images in parallel

Benchmarks

dssim runs single-threaded here, so both sides use one core. dssim is a different algorithm, so the chart compares cost, not scores. Decode is excluded.

Rust (`blazediff-ssim` vs `dssim`) (image IO excluded)

TOTAL TIME, BEST-OF · 16 FIXTURES · IO EXCLUDED

BlazeDiffOther librariesLower is better

dssim is a different algorithm (multi-scale, Lab chroma), so this compares cost, not scores.

Apple M1 Max · darwin/arm64 · Node v22.14.0 · rustc 1.97.1 (8bab26f4f 2026-07-14) · best-of, iterations scaled by image size, dssim single-threaded · 2026-10-22 · 234b300-dirty · pnpm benchmark ssim-dssim

Per-fixture results (16)

Fixture	MPx	dssim	ssim	ms-ssim	hitchhikers	ssim vs dssim	ms-ssim vs dssim	hitchhikers vs dssim
4k/1	17.9	1.78s	99.32ms	250ms	131ms	17.93x	7.13x	13.63x
4k/2	20.0	1.80s	117ms	279ms	162ms	15.35x	6.45x	11.12x
4k/3	24.0	2.31s	159ms	331ms	172ms	14.55x	6.98x	13.46x
alpha/1	0.0	4.10ms	0.47ms	0.75ms	0.23ms	8.74x	5.46x	17.67x
blazediff/1	0.4	31.47ms	5.15ms	5.89ms	2.88ms	6.11x	5.34x	10.91x
blazediff/2	0.4	26.45ms	1.54ms	4.95ms	2.29ms	17.16x	5.35x	11.58x
blazediff/3	1.6	113ms	5.07ms	22.99ms	11.21ms	22.24x	4.90x	10.05x
blazediff/4	3.8	252ms	12.28ms	52.49ms	24.92ms	20.53x	4.80x	10.12x
pixelmatch/1	0.1	11.75ms	1.54ms	2.10ms	0.82ms	7.65x	5.60x	14.27x
pixelmatch/2	0.1	6.89ms	0.76ms	1.21ms	0.42ms	9.08x	5.67x	16.57x
pixelmatch/3	0.1	11.61ms	1.59ms	2.19ms	0.87ms	7.28x	5.29x	13.30x
pixelmatch/4	0.2	15.28ms	0.78ms	2.60ms	1.17ms	19.50x	5.87x	13.03x
pixelmatch/5	0.1	6.70ms	0.76ms	1.21ms	0.41ms	8.84x	5.53x	16.19x
pixelmatch/6	0.1	6.68ms	0.75ms	1.21ms	0.41ms	8.95x	5.52x	16.13x
pixelmatch/7	0.3	19.53ms	1.07ms	3.56ms	1.64ms	18.27x	5.48x	11.89x
same/1	1.7	120ms	5.76ms	24.75ms	11.50ms	20.86x	4.85x	10.45x
TOTAL		6.51s	413ms	986ms	523ms	15.78x	6.61x	12.46x

Switching


use blazediff_ssim::{ms_ssim, MsSsimOptions, Plane, Rgba8, SsimOptions};
 
let base = Plane::from_rgba8(Rgba8::new(&rgba1, width, height))?;
let current = Plane::from_rgba8(Rgba8::new(&rgba2, width, height))?;
 
let outcome = ms_ssim(&base, &current, &SsimOptions::default(), &MsSsimOptions::default())?;
println!("{:.6}", outcome.score); // 1.0 means identical

A dssim threshold does not convert to an SSIM threshold. Score your own fixtures with the new metric and pick the threshold from those scores.

When to keep dssim

You tune image codecs or compression settings. dssim was validated on TID2013 for that job.
You need color-profile handling, or errors that only show in chroma.
Your project is already AGPL, or you bought the commercial license.