What is VMAF?
VMAF is a full-reference perceptual metric developed by Netflix with academic collaborators to predict subjective video quality from reference and distorted sequences. It combines multiple image-quality features into a score.
How VMAF works
VMAF evaluates an impaired sequence against a temporally and spatially corresponding source, extracts several perceptual features, and fuses them through a trained model. This makes it more viewing-oriented than a single pixel-error formula, but the result still reflects the selected model, preprocessing, and pooling method. Encoding pipelines use it for compression and scaling experiments, regression tests, and ladder analysis, then confirm important decisions with scene inspection and playback evidence. Results should not be generalized to packet loss, corruption, or other artifact classes absent from a model’s training data.
Key facts
- Reference and distorted frames must represent the same moments and comparable image geometry; offsets, dropped frames, or inconsistent scaling can measure misalignment instead of degradation.
- A VMAF result is tied to a particular model and tool configuration. Scores produced with different models or preprocessing settings should not be assumed directly comparable.
- A single pooled score can conceal a short, severely damaged scene. Per-frame traces and lower-tail or scene-level summaries help expose localized quality failures.
When VMAF matters
Encoding teams calculate VMAF when comparing codecs, scaling choices, encoder settings, or bitrate ladders. Results depend on the selected model, reference, and artifact types represented in that model’s training data, so one score should not replace playback testing.