Block 2. Objective Metrics

PSNR measures pixel error in decibels: fast, universal, but a weak predictor of how video looks. See the formula, the dB scale, and where PSNR misleads.

SSIM measures whether a video's structure survived, not just pixel error — so it beats PSNR. See the formula, a worked example, and where SSIM still lies.

MS-SSIM measures video structure across five scales, not one, so it beats single-scale SSIM across screen sizes. The weights, a worked example, and FFmpeg.

VMAF is Netflix's perception-trained video quality metric on a 0–100 scale. See what it fuses, the 2026 v1 models, the 6-point JND, and where it lies.

A VMAF score depends on the model, pooling, and error bar. See how to pick the model, read the confidence interval, and pool without fooling yourself.

VMAF-NEG stops sharpening and contrast tricks from inflating a VMAF score. See what it changes, the gaming numbers, and when to report NEG vs the default.

Beyond VMAF: how ITU-T P.1204 scores quality with no reference, what Apple's AVQT adds, and the learned metrics for AI codecs — and whether you need them.

How video quality metrics are validated against human MOS — what PCC, SROCC, and RMSE measure, the test databases, and why correlation may not transfer.

How per-frame VMAF and SSIM scores are pooled into one number — why the arithmetic mean hides bad frames, and when to use harmonic, percentile, or min.

Where PSNR, SSIM and VMAF mislead — film grain, motion, dark scenes, banding, chroma, screen content — which metrics fail and what to measure instead.

PSNR, SSIM, or VMAF? Match the metric to the job — regression gate, codec comparison, delivered quality, or live — with a decision tree and free tool.

How to read a VMAF or SSIM report like a skeptic: check the model, pooling, confidence interval, resolution, and reference before you trust the number.