VMAF, short for Video Multimethod Assessment Fusion, is a full-reference video quality metric that predicts how viewers would rate a compressed or scaled video compared with its pristine source. In video quality work, VMAF has become the most widely used perceptual metric for tuning encoders, building bitrate ladders and comparing codecs, because it correlates with subjective scores far better than PSNR on typical streaming content.
The metric was developed by Netflix together with university researchers and released as open source in 2016. It reports a score from 0 to 100, where higher is better, and it is computed frame by frame and then pooled into one value for a clip.
How VMAF fuses quality features
VMAF does not measure quality directly. It computes several elementary features that each capture one kind of degradation, and a trained regression model combines them into a single prediction of the mean opinion score.
The default model uses Visual Information Fidelity at four scales, a detail loss metric that separates lost detail from added impairments, and a temporal motion feature based on differences between neighboring frames. A support vector regression model, trained on viewer ratings collected under controlled conditions, maps these features to the final score.
| Feature | What it measures | Degradations it reflects |
|---|---|---|
| VIF (four scales) | Information preserved relative to the source | Blur, noise, compression loss |
| DLM or ADM | Loss of detail versus added distortion | Smearing, ringing, blocking |
| Motion | Average luma difference between frames | Masking of artifacts in fast motion |
| SVR fusion | Weighted combination of all features | Overall perceived quality |
Because the model is trained, VMAF inherits the viewing conditions of its training data. The standard model assumes a 1080p display viewed at about three times the screen height, so the same encode can deserve a different score on a phone or a large 4K television.
VMAF models and score ranges
Several official models exist for different viewing situations, and choosing the wrong one is a common source of misleading results. Each model also has a NEG variant, meaning no enhancement gain, which penalizes sharpening and contrast boosts that could otherwise inflate scores.
| Model | Viewing assumption | Typical use |
|---|---|---|
| vmaf_v0.6.1 | 1080p display, 3H viewing distance | Default for HD streaming |
| Phone model | Small mobile screen | Mobile-first delivery decisions |
| vmaf_4k_v0.6.1 | 4K display, 1.5H viewing distance | UHD encoding ladders |
| NEG models | Same as base model | Codec comparisons, preprocessing checks |
Scores are interpreted in bands rather than absolute thresholds. The values below are typical working ranges in streaming practice for the default model, not fixed limits, and they shift with content and display.
| Score range | Typical perception | Common use |
|---|---|---|
| 93 to 100 | Hard to distinguish from the source | Top rendition of a ladder |
| 80 to 93 | Good, minor artifacts on close inspection | Mid-ladder renditions |
| 60 to 80 | Visible artifacts, acceptable on small screens | Low-bandwidth renditions |
| Below 60 | Clearly degraded | Emergency or very low bit rates |
A difference of about six VMAF points is often treated as roughly one just noticeable difference, but this rule of thumb depends on content and model. Differences of one or two points are usually within measurement noise for a single clip.
VMAF compared with PSNR and SSIM
PSNR, SSIM and VMAF are all full-reference metrics, so they need the original source aligned frame by frame with the test video. They differ in what they model and how closely they track human judgment.
| Metric | Basis | Strengths | Weaknesses |
|---|---|---|---|
| PSNR | Mean squared pixel error | Simple, fast, mathematically clear | Poor match with perceived quality |
| SSIM and MS-SSIM | Local structure, contrast, luminance | Better perceptual match than PSNR | Weak on temporal effects |
| VMAF | Fused features with trained model | Strong correlation on streaming content | Luma only, can be gamed by sharpening |
In practice, engineers report VMAF together with PSNR or SSIM. When the metrics disagree strongly, the disagreement itself is a signal to inspect the frames visually.
Using VMAF correctly
VMAF is reliable only when the comparison is set up correctly. The distorted video must be scaled back to the reference resolution, matched in frame rate and aligned exactly in time; a single dropped frame can collapse the score.
| Mistake | Effect on the score | Correct practice |
|---|---|---|
| Misaligned frames | Large, unrealistic drop | Check frame counts and offsets first |
| Wrong upscaling filter | Bias between renditions | Use the same scaler, such as bicubic, for all |
| Mean pooling only | Short bad sections hidden | Also inspect low percentiles and minimum |
| Wrong model for display | Scores that do not match viewing | Choose phone, HD or 4K model deliberately |
| Sharpening before encode | Inflated scores | Report NEG model alongside default |
Pooling deserves attention. The arithmetic mean can hide a few seconds of severe artifacts, so many teams also track the harmonic mean or the 5th percentile of frame scores to catch short quality drops that viewers notice.
Limitations of VMAF
VMAF is a model of average viewer opinion, not a ground truth. It was trained mainly on compression and scaling artifacts in SDR content, so it is less dependable outside that domain.
- Color is not evaluated; the standard model uses luma only, so chroma subsampling errors and color shifts go unnoticed.
- HDR content requires care, because the standard model was not trained on PQ or HLG signals.
- Banding in smooth gradients is often underestimated compared with how visible it is to viewers.
- Film grain synthesis or removal can produce scores that contradict visual judgment.
Put simply
VMAF is a computer’s estimate of how good a video looks to people. It compares a compressed video with the original and gives a score from 0 to 100, where 100 means it looks the same.
Older measures only counted how many pixels changed. VMAF instead was taught by real viewers’ ratings, so it cares more about the kinds of damage people actually notice, such as blur and blocky patches.
In practice
An online learning platform wants to cut its delivery costs without making lectures look worse. The video team encodes a set of representative clips at several bit rates for each resolution and measures VMAF with the default model, scaling every rendition back to 1080p with the same bicubic filter.
They find that the 1080p rendition reaches a mean VMAF of 94 at 3 Mbps and only 95 at 5 Mbps for slides and talking-head footage, while the 5th percentile stays above 88 at both. They lower the top rendition to 3 Mbps, confirm with a short visual review and keep the higher rate only for clips with complex motion. The result is a noticeably smaller bandwidth bill with no visible quality loss reported by viewers.