VMAF — Video Multi-method Assessment Fusion — is a quality metric developed and open-sourced by Netflix. It compares a compressed video frame by frame against the original and produces a score from 0 to 100, trained to track what human viewers actually report rather than what raw pixel differences suggest.
Rough reading of the scale:
- 95 and above — differences are very hard to see in normal viewing.
- 90 to 95 — visible if you look for them, on the hard frames.
- Below 85 — visible without looking for them.
- Below 75 — obviously degraded.
Two caveats that matter when reading anyone's numbers, including ours. First, VMAF is a mean across frames, and a mean hides its worst moments; a clip averaging 93 can contain a second of mush. The minimum per-frame score is the more honest figure for a hard cut, and our sweep records both. Second, VMAF was trained on television-style content at viewing distances to match, so it is least reliable on the content types furthest from that — heavy screen text and synthetic animation among them.
What it is genuinely good for is comparison under fixed conditions: same source, same encoder, one variable changed. That is the only claim made with it here. A VMAF score quoted with no source clip, no encoder and no preset is not a measurement, it is a decoration.