Does VMAF compress video?

No. It evaluates video quality and can inform the choices made by a separate encoder.

Does a score of 90 mean 90% quality?

No. Interpret it using the selected model's scale and evaluation conditions.

Can the same video receive different scores with different models?

Yes. Models can represent different training data, features and viewing assumptions.

AI Glossary / MLOps and deployment / Beginner

Video Multimethod Assessment Fusion

Definition

Video Multimethod Assessment Fusion, usually written Video Multi-Method Assessment Fusion or VMAF, is a perceptual video quality assessment method developed by Netflix. It combines measured visual features through a trained model to estimate perceived quality, normally by comparing a processed video with a reference video under defined evaluation conditions.

What it means

A compressed video can differ from its source in many ways: detail may disappear, edges may change, or visible artifacts may appear. Counting pixel differences alone does not necessarily describe how noticeable those changes are to viewers. VMAF uses a learned relationship between visual measurements and human quality judgments to produce a more perceptually informed estimate.

The score is a model prediction, not a direct reading of every viewer's experience. A meaningful result includes the model, video preparation and viewing assumptions. It does not certify that a video is accurate, accessible, enjoyable or suitable for an AI recognition task.

How it works

The evaluator prepares corresponding reference and processed pictures so that the comparison represents the intended presentation. It extracts visual measurements and supplies them to a trained prediction model. Scores can be examined over time or combined into a summary through a chosen pooling method.

Models differ in training, features and intended viewing conditions. Netflix's documentation includes models for different displays and viewing distances. Consequently, a score without a model identifier is incomplete evidence, especially when comparing results generated by separate teams.

The reference also matters. Comparing against an already damaged source assesses changes relative to that source; it does not establish an absolute ideal. Misaligned frames, inconsistent scaling or incorrect pixel interpretation can make a technically successful calculation answer the wrong question. Store these preparation choices with the reported result.

Three practical examples

1. Course delivery: a publisher compares several encodes of the same lesson, balancing data size against estimated visual quality. It also checks whether fine writing on slides remains readable.

2. Encoder evaluation: engineers compare rate-quality curves across a representative collection of clips. They hold evaluation settings constant and inspect differences that might come from preprocessing rather than compression.

3. AI video pipeline: a team uses VMAF as one image-quality indicator while separately measuring object-detection accuracy. It does not assume that a good perceptual score guarantees good machine-analysis results.

Think of it this way

Think of a trained tasting panel's scoring guide converted into a measuring instrument. Several measurements jointly predict how the panel might rate a prepared sample, and the instrument works best within the conditions represented during calibration.

The analogy explains combining evidence and learning from people. VMAF does not taste, feel or watch with human awareness. Its measurements and predictions can miss distortions or situations that matter to a particular viewer.

Worked example

Suppose two ten-minute encodes are measured with the same model and preparation procedure. Encode A uses 4 megabits per second and receives an illustrative mean score of 92. Encode B uses 3 megabits per second and receives 90.

Ignoring audio and container overhead, A contains 4 × 600 ÷ 8 = 300 megabytes of video. B contains 3 × 600 ÷ 8 = 225 megabytes. B therefore saves 75 megabytes, or 25% of A's video size.

The quality prediction falls by two score points. This is not proof that quality falls by 2%, nor that every viewer notices the difference. Before selecting B, inspect its lowest-quality scenes and conduct relevant viewing checks. These invented scores illustrate decision arithmetic; they are not measured codec results or a universal acceptable threshold.

Illustrative case study

Fictional case: ClearLesson tests a new encoding recipe for educational videos. Its mean VMAF result looks acceptable across the evaluation set, but instructors report difficulty reading a small equation during one short scene.

The team inspects the score timeline and the actual scene. It expands its review to include text-heavy material and records readability separately from the aggregate quality estimate. It also keeps the exact model file and scaling procedure in its evaluation record so later comparisons remain reproducible.

ClearLesson's lesson is to use the metric as evidence within a content-specific decision. The case does not assert that VMAF always misses text or that one additional metric solves every evaluation problem.

Common misconceptions and mistakes

Do not interpret a score as a percentage of original pixels preserved. It is a perceptual prediction on the chosen model's scale.

Do not assume every VMAF model always uses an identical 0-100 range. Netflix's documented model variants include differing operating ranges, so consult the selected model's definition.

Do not compare scores produced under inconsistent scaling, alignment or model settings as though only the encoder changed. For codec comparisons, examine whether enhancement operations affect the result and whether an appropriate No Enhancement Gain configuration is required.

Business use

VMAF can help teams make encoding decisions consistently across large media collections. Its usefulness grows when it is combined with bitrate, processing cost, viewing checks and playback measurements.

Define evaluation content before choosing a target, including motion, texture and instructional detail relevant to the service. Record exceptions as well as averages. Buffering, startup delay, audio and accessibility need their own assessment; a perceptual video score cannot stand in for the entire customer experience.

Check your understanding

In the worked example, what does the 25% figure establish?

Answer: It establishes the reduction in video data size under the stated bitrate and duration assumptions. It does not establish the percentage change in perceived quality or viewer satisfaction.

Related terms

Full-reference metric
Video codec
Model validation
Perceptual quality
bitrate
PSNR
SSIM
subjective evaluation
temporal pooling
rate-quality curve

Learn next

Learn about subjective video assessment and rate-quality curves next. Compare several encodes of one short reference clip using identical settings, then inspect difficult scenes directly. Record the model identifier and preparation steps alongside every score.

Sources

Netflix, VMAF project and documentation: https://github.com/Netflix/vmaf

Netflix, VMAF: The Journey Continues, methodology and evaluation practices: https://netflixtechblog.com/vmaf-the-journey-continues-44b51ee9ed12

Netflix, model selection and operating ranges: https://github.com/Netflix/vmaf/blob/master/resource/doc/models_v1.md

Netflix, earlier models and No Enhancement Gain: https://github.com/Netflix/vmaf/blob/master/resource/doc/models_v0.md

Primary sources consulted. Model details are version-dependent; numerical examples and the case study are illustrative.

Previous: Video magnification Next: Vietoris - Rips filtration