What is a good SSIM value?

SSIM runs from −1 to 1, where 1 means the images are identical. As a rule of thumb, 0.99 and above is essentially identical, 0.95–0.99 is excellent, 0.90–0.95 is good, 0.80–0.90 is fair, and below 0.80 shows noticeable structural distortion. These bands are an engineering convention, not a formal standard.

How is SSIM calculated from two images?

Both images are converted to luminance, then an 11×11 Gaussian window (σ=1.5) slides across them. Each window's SSIM combines a luminance, contrast and structure term using local means, variances and covariance. Averaging the per-window scores gives the Mean SSIM (MSSIM) reported for the whole image.

What is the difference between SSIM and PSNR?

PSNR measures pixel-level error on a decibel scale and is fast but does not track how the eye perceives quality. SSIM compares structure, luminance and contrast in local windows and usually correlates better with perceived quality. A blurry image can keep a fair PSNR while its SSIM drops. Many papers report both.

What does an SSIM of 1 mean?

An SSIM of exactly 1 means the two signals are identical at every pixel — the luminance, contrast and structure terms all equal 1, so their product is 1. Any difference, even a uniform brightness shift, pushes the score below 1. A value of 0 means no structural similarity, and negative values mean the structure is anti-correlated.

What are C1 and C2 in the SSIM formula?

C1 and C2 are small stabilising constants that stop the formula dividing by near-zero when local means or variances are tiny. They are defined as C1 = (K1·L)² and C2 = (K2·L)², where L is the dynamic range (255 for 8-bit) and K1 = 0.01, K2 = 0.03 by default. At L = 255 that gives C1 = 6.5025 and C2 = 58.5225.

Why does a brightness shift lower SSIM but not the structure?

Adding a constant to every pixel changes the mean but leaves variances and covariance untouched. So the contrast and structure terms stay near 1 while the luminance term drops. SSIM therefore registers the change, unlike a pure variance-based measure. This is one reason SSIM tracks perceived quality more closely than raw error.

Does this SSIM match scikit-image?

Yes. Image mode uses Gaussian weights (σ=1.5), the sample-covariance normalisation cov_norm = NP/(NP−1), and the same border crop as scikit-image's structural_similarity with data_range set to L. For a given pair it agrees within about ±0.01. Verified on 2026-06-26.

Is my image uploaded anywhere?

No. The image is decoded and analysed entirely in your browser using the Canvas API. Nothing is sent to a server, there is no signup, and no copy is stored. You can confirm this in your browser's network tab — there are no upload requests when you run the calculation.

Developer · Image & Video

SSIM Calculator — Structural Similarity Index

Measure how perceptually close two images are. Upload a reference and a distorted version to get the Mean SSIM over an 11×11 Gaussian window, or paste two pixel lists to hand-check the maths. You get the score, a quality band, and the luminance, contrast and structure breakdown. Verified against scikit-image.

By Induwara Ashinsana— Executive Director, Ryzera TechnologiesUpdated Jun 26, 2026

Calculate SSIM

Wang et al. · scikit-image verified

Constants

Dynamic range L

C1 = (K1·L)² = 6.5025 · C2 = (K2·L)² = 58.5225

Reference values

Comma-, space-, or newline-separated. Each value 0–255.

Distorted values

Must have the same number of values as the reference list.

SSIM

0.9845

Samples compared

Luminance · Contrast · Structure

1.0000 · 0.9882 · 0.9963

Quality band

Excellent

Very close to the reference; high quality.

Term	Inputs	Value
Luminance l(x,y)	μx=25 · μy=25	1.0000
Contrast c(x,y)	σx²=166.6667 · σy²=118.6667	0.9882
Structure s(x,y)	σxy=140	0.9963
SSIM = l · c · s		0.9845

Cross-check

Component product l·c·s = 0.9845 matches the combined single-fraction form = 0.9845.

All maths runs in your browser — no image ever leaves your device. Formula and constants from ; image mode is cross-checked against scikit-image.

How it works

SSIM (Structural Similarity Index) measures how close a distorted image — compressed, denoised, super-resolved or generated — is to a reference, but unlike pixel-error metrics it models how the human eye reads a scene: as structure rather than isolated pixels. It is defined by Wang, Bovik, Sheikh & Simoncelli (2004) and implemented as the de-facto standard by scikit-image, which this tool cross-checks against.

Local statistics. For two aligned signals x and y over a window, compute the means μx, μy, the sample variances σx², σy² (divided by N−1 to match scikit-image), and the covariance σxy.
Stabilising constants. C1 = (K1·L)² and C2 = (K2·L)² with K1 = 0.01, K2 = 0.03 and L the dynamic range (255 for 8-bit) — giving C1 = 6.5025 and C2 = 58.5225 at L = 255.
Three comparison terms. Luminance l = (2μxμy + C1)/(μx² + μy² + C1), contrast c = (2σxσy + C2)/(σx² + σy² + C2), and structure s = (σxy + C2/2)/(σxσy + C2/2).
Combine. With the usual weights (α=β=γ=1), SSIM = l · c · s, which is algebraically the single fraction [(2μxμy + C1)(2σxy + C2)] / [(μx² + μy² + C1)(σx² + σy² + C2)]. The tool shows both and confirms they agree as a cross-check.
Mean SSIM for whole images. Both images are converted to luminance, then an 11×11 Gaussian window (σ=1.5) slides across them computing a local SSIM at each position. Averaging those windows gives the Mean SSIM (MSSIM). Identical dimensions are required and validated up front; nothing is uploaded — all of this runs in your browser.

The coloured quality band (essentially identical ≥ 0.99, excellent 0.95–0.99, good 0.90–0.95, fair 0.80–0.90, poor < 0.80) is a widely used heuristic for a quick read, not part of any standard — the page says so next to the result. SSIM is single-scale here; multi-scale MS-SSIM and video SSIM are out of scope for this version.

Worked examples

Short pixel lists (hand-checkable)

Reference [10, 20, 30, 40] vs Distorted [12, 22, 28, 38], L = 255

Means: μx = 25, μy = 25
Sample variances (÷3): σx² = 500/3 ≈ 166.667, σy² = 356/3 ≈ 118.667
Covariance: σxy = 420/3 = 140
l = (2·625 + 6.5025)/(1250 + 6.5025) = 1.0000
c = (2·12.910·10.893 + 58.5225)/(285.333 + 58.5225) = 0.98817
s = (140 + 29.261)/(140.626 + 29.261) = 0.99627
SSIM = l·c·s = 0.9845 → band: excellent

Pure brightness shift (edge case)

Distorted = Reference + 50: [10,20,30,40] vs [60,70,80,90], L = 255

Means: μx = 25, μy = 75 (variances and covariance unchanged)
l = (2·25·75 + 6.5025)/(625 + 5625 + 6.5025) = 0.60042
c = 1.0000 and s = 1.0000 (structure is identical)
SSIM = 0.6004 → band: poor
PSNR-from-MSE would only see the squared error; SSIM flags the shift.

Identical images

Reference = Distorted → no distortion

μx = μy, σx² = σy² = σxy at every window
Each of l, c and s evaluates to exactly 1
SSIM = 1·1·1 = 1.0000
Image mode returns Mean SSIM = 1.0000 → band: essentially identical

Frequently asked questions

Sources & references

The SSIM formula, the C1 = (K1·L)² / C2 = (K2·L)² constants, and the 11² Gaussian-window Mean SSIM on this page were last cross-checked against the Wang et al. (2004) paper and scikit-image on 2026-06-26. The quality bands are explicitly labelled as heuristics, not a standard.

Related tools

LiveDeveloper

PSNR Calculator

Calculate Peak Signal-to-Noise Ratio (PSNR) in decibels for image and video quality — from a known MSE and bit depth, or from two pixel-value lists. Shows MAX, the MSE used, a quality band, and the full derivation. Free, no signup.

Open tool

LiveDeveloper

YAML ⇄ JSON Converter

Convert YAML to JSON and JSON to YAML entirely in your browser. Auto-detects direction, expands anchors/aliases and merge keys, handles multi-document streams, locates parse errors by line, and verifies each conversion with a round-trip check. No upload, no signup.

Open tool

LiveDeveloper

JavaScript Playground

Write and run JavaScript online, free, no signup. Code runs in a sandboxed worker in your own browser with live console output.

Open tool

Rate this tool

Be the first to rate

Comments & feedback

Spotted a bug or want an improvement? Tell us — our team reviews every comment, and good ideas get built. Comments are public and anonymous.

Found a bug, edge case, or want to suggest an improvement?

Email me at [email protected] — most fixes ship within 24 hours.