SSIM — Structural Similarity Index — compares local patterns of luminance, contrast and structure between a compressed frame and its original. It returns a value from 0 to 1, where 1 is identical. It predates VMAF by about a decade and is much cheaper to compute.
Its weakness shows up clearly in our own CRF sweep, which recorded both metrics on the same encodes. On camera footage at CRF 32 — visibly degraded, VMAF 74.4 — SSIM still read 0.973. To three decimal places that looks like a near- perfect reproduction. It is not; the metric is simply not very sensitive to the kinds of loss that bother people at that end of the scale.
The inverse is also visible: on screen recordings, SSIM sits above 0.996 across every CRF value we tested, from 18 to 32, while VMAF moves a meaningful 6.9 points over the same range. A metric that barely changes while the picture changes is not measuring what you want to know.
So SSIM is kept here as a secondary reading, not a headline. It is genuinely useful for confirming that two encodes are structurally the same file rather than accidentally different sources, and it is worth being sceptical of any comparison that quotes a high SSIM without a perceptual metric beside it — 0.98 sounds like a guarantee and is not one.