VMAF Video Quality Metric

VMAF predicts perceived video quality by comparing a decoded test video with a reference under a selected model and viewing context.

Suppose you are making smaller playback versions of the same retained movie source. Looking at every frame of every candidate is impractical, but file size alone says little about which version preserved the picture better. VMAF supplies one way to measure that comparison.

VMAF, Video Multi-Method Assessment Fusion, is a full-reference video-quality metric. It receives decoded pictures from a reference and a test sequence, extracts comparison features, and uses a trained model to predict quality in a defined viewing context.

What goes into the score

The comparison begins with corresponding pictures, not with file extensions or compressed byte sizes. Features capture aspects of the relationship between the reference and test content, such as retained detail and motion-related behavior. A model combines the features using relationships learned from subjective-quality data.

The published v0.6.1 model, for example, lists an ADM detail feature, a motion feature, and VIF features at four spatial scales. These give the model several measurements of the picture relationship rather than one overall sample-error value. The selected model defines which measurements it consumes.

Training supplies examples for which comparison features and viewer assessments are both available. The fitting process learns a mapping from the feature combinations to those assessments. Evaluation then measures a new reference/test pair and applies the fitted mapping. The training happened before this evaluation; choosing a VMAF model does not train a new model on the movie being tested.

That last stage is why the result differs from a direct error formula such as PSNR. VMAF's number is the output of a particular learned model. It is intended to be useful for perceptual comparison; it is not a count of retained bits, pixels, or original information.

A score of 95 therefore needs context. Which model produced it? How were the inputs prepared? Is it a mean for a whole film, or the score of one scene?

Prepare the comparison before interpreting it

Corresponding input pictures must represent the same content at the intended comparison geometry. A frame offset can compare two different moments. A crop difference can compare different regions. A range or color conversion applied on only one side can become part of the measured difference.

When comparing a lower-resolution playback version with a larger reference, scaling is an explicit part of the evaluation. Use a consistent preparation that represents the intended viewing situation. A different scaler can change what the metric sees.

This is especially relevant to HDR-to-SDR material. A straightforward comparison of incompatible signal representations does not isolate compression quality. Establish which picture representation is being compared and whether the selected model is suitable for it.

For repeatable results, keep a small record with each test:

  • The reference source and the tested encode.
  • Model/version and tool settings.
  • Sample format, geometry, scaling, and timing preparation.
  • How per-frame results were aggregated.

Viewing distance and model-specific score scales

Artifacts that are visible close to a large screen may be much less visible from farther away. Models calibrated for different viewing conditions can therefore assign different values to the same reconstruction.

Netflix's familiar v0 models and the v1 models released in June 2026 have documented viewing assumptions. Common models use a 0–100 scale; the v1 4K model for viewing at three screen heights uses 0–110. These are model conventions, not a universal percentage scale.

A score from one model should not be treated as the same quantity as an equal-looking score from another. Even within one model, changing clipping settings or input preparation can change the interpretation of the result.

Look past the whole-film average

Consider a hypothetical constant-frame-rate, one-minute test: fifty seconds score 95, and ten seconds score 50. The arithmetic mean is 87.5. That single number cannot tell you whether the lower-scoring section is an inconsequential transition or the scene the viewer most cares about.

Per-frame and per-scene results help locate the weakness. A low-tail summary can be useful alongside the mean, and a targeted viewing check can show what happened: smeared grain, unstable texture, blurred fine detail, or another visible change.

Enhancement processing can complicate the measurement. Sharpening may change measured features while introducing a look that a viewer dislikes. Grain handling can also affect a reference comparison even where a synthetic-grain presentation is intentional. The metric and a viewing check answer related questions, rather than replacing one another.

For an Emby library conversion, use VMAF to narrow the candidate encodes and identify passages worth inspecting. After choosing a satisfactory picture reconstruction, test the intended playback path separately for decoding and delivery performance. That preserves a clear meaning for the score: how the tested pictures compare under the stated model and preparation.

See Also

References