Learning course · Updated June 2026
How delivered video quality is actually measured — the discipline one step downstream of encoding and streaming. PSNR, SSIM, and VMAF from first principles, subjective testing that survives scrutiny, a labelled artifact gallery, production QC gates, streaming QoE, and Fora Soft's own benchmarks on real content. A vendor-neutral, measurement-honest course from Fora Soft engineers.
Every metric has a blind spot, and we name it. Every claim is tied to a named source and year — the SSIM paper (Wang et al., 2004), the Netflix VMAF docs, ITU-T P.910 / P.808 / P.1204, ITU-R BT.500, and Bjøntegaard's BD-rate. A single number is a summary, not the truth — this course teaches you to read it honestly.
Outcomes
Eight blocks that take you from "the picture looks fine to me" to a defensible quality number. By the end, you can choose the right metric, run a subjective test that holds up, diagnose any artifact, gate a pipeline on quality, and read a benchmark without fooling yourself.
Pick a path
The same 57 articles, ordered for what you actually need to do this quarter.
Why a number beats an opinion, and the metrics that produce it. QoE vs QoS, the reference taxonomy, then PSNR, SSIM, and VMAF explained in full — math once, worked example, blind spots named.
The practitioner's middle. Run a human-rating test that survives scrutiny, identify any artifact on sight and trace it to its cause, then turn quality into an automated CI gate.
The operator's edge. Measure the viewer's experience, read and reproduce a codec benchmark with BD-rate, and wire the whole thing up with FFmpeg and libvmaf.
Syllabus
Every chapter is self-contained. Read in order, or jump straight to the block you need — from why we measure to the tools that do it.
01
02
03
04
05
06
07
08
Fora Soft validates every streaming and encoding decision against quality numbers — PSNR, SSIM, VMAF, and our own benchmarks since 2005.
Featured
Hand-picked deep dives — the three metric anchors everyone searches, plus the benchmark work that makes this section worth citing.
Reference
120+ terms with crisp, cited definitions, aliases, and links to the deep dives. From PSNR, SSIM, and VMAF to MOS, BD-rate, banding, and the convex hull — the full A–Z of video-quality measurement is one click away.
PSNR
Peak Signal-to-Noise Ratio. The baseline full-reference metric: the ratio of maximum signal to the error versus a reference, in decibels. Simple and fast, but a weak match for human perception — the metric every comparison starts from.
SSIM
Structural Similarity Index (Wang et al., 2004). Compares luminance, contrast, and structure in a sliding window instead of pixel error, so it tracks perceived quality better than PSNR. Scored 0–1.
VMAF
Video Multi-Method Assessment Fusion (Netflix). A machine-learned metric that fuses several quality features and is trained against human scores; the de-facto modern metric, with models for phone, 4K, and the no-enhancement-gain VMAF-NEG variant.
MOS
Mean Opinion Score. The average of human quality ratings on a fixed scale (usually 1–5) — the ground truth every objective metric is validated against. DMOS is its differential form.
BD-rate
Bjøntegaard Delta rate. The standard way to express how much bitrate one encoder or codec saves another at equal quality — a single percentage that summarizes two rate-quality curves.
Banding
The visible stair-stepping in what should be a smooth gradient (sky, shadow), caused by too few code values — a classic compression artifact that PSNR and SSIM often miss but viewers always see.
Written and maintained by
FAQ
Video quality measurement assigns a defensible number to how good a video looks, instead of relying on opinion. It splits into objective metrics — algorithms like PSNR, SSIM, and VMAF that compare a processed video to a reference — and subjective testing, where humans rate quality to produce a Mean Opinion Score. The objective metrics are validated against the human scores. It sits downstream of encoding and streaming, and it is how teams prove a change helped, not hurt.
PSNR (Peak Signal-to-Noise Ratio) is the baseline full-reference metric. It expresses, in decibels, the ratio between the maximum signal and the mean-squared error versus a reference — higher is better, and values above roughly 40 dB usually look good. It is simple, fast, and universally supported, so every comparison starts with it, but it correlates only loosely with perception and misses artifacts like banding — rarely the metric you finish with.
All three are full-reference picture-quality metrics of increasing sophistication. PSNR measures raw pixel error in decibels — fast but a weak match for the eye. SSIM compares luminance, contrast, and structure in a sliding window, tracking perception better, scored 0 to 1. VMAF is machine-learned, fusing several features and trained against human ratings, so it usually correlates best. Use PSNR for a sanity check, SSIM for structure, VMAF for a perception-aligned score — and know each blind spot.
VMAF-NEG (No Enhancement Gain) is a VMAF variant designed to resist gaming. Standard VMAF can be inflated by sharpening or contrast tricks that raise the score without improving fidelity — output that looks better than the reference rather than closer to it. VMAF-NEG removes that gain, reporting how faithfully the output matches the source. Use it when comparing encoders or settings and you need a score a post-filter cannot juice.
BD-rate (Bjontegaard Delta rate) is the standard way to express how much bitrate one codec or encoder saves another at equal quality. Instead of comparing single points, it integrates the area between two rate-quality curves into one percentage — for example, AV1 giving 30% BD-rate savings over H.264 means equal quality at 30% less bitrate. It can be computed against PSNR, SSIM, or VMAF, and it is the headline number in every serious codec comparison.
FFmpeg computes PSNR and SSIM directly with its libavfilter filters, and VMAF through the libvmaf filter (built with the VMAF library). You pass the processed video and the reference, make sure they are aligned and at the same resolution and frame rate, choose the right VMAF model, and read the pooled score from the log. The common pitfalls are mismatched scaling, frame misalignment, and the wrong model — get those right and FFmpeg is the everyday workhorse.
Fora Soft has built real-time video, audio, and AI products since 2005 — WebRTC, LiveKit, generative pipelines, and AI agents at scale. Tell us what you’re building and we’ll send a real engineer your way.
Specialist software house for video, real-time and AI products. Founded 2005. 50 in-house engineers.
