Smaller, but how much worse? What VMAF tells you
VMAF is Netflix's open-source video quality metric, scored 0 to 100. What the numbers mean, what score to aim for when you compress, and where VMAF falls short.
VMAF (Video Multi-method Assessment Fusion) is an open-source quality metric from Netflix that compares a compressed video with its original and predicts, on a 0–100 scale, how a viewer would rate the difference. Higher is better, and a score in the mid-90s usually means the difference is very hard to see in normal viewing.
“90% smaller with no quality loss” is the most common claim in video compression, and it’s almost never measured. That’s why every file Becheloon makes comes with a VMAF number instead of a promise.
How does VMAF work?
VMAF reads the original and the new file frame by frame and runs several simpler measurements on them: how much fine detail survived, how well edges and structure held up, and how much motion there is. A machine-learning model, trained on ratings that real people gave to real test clips, combines those into one score. Netflix developed it with university researchers, released it as open source in 2016, and it can be run through FFmpeg.
What is a good VMAF score?
There’s no official pass mark. This rough guide matches how the score is commonly read:
| VMAF | What it usually means |
|---|---|
| 95 and above | Differences very hard to see in normal viewing |
| 90–95 | Fine for social feeds on phones; a trained eye on a big screen may notice |
| 85–90 | Softness or artefacts start to show in demanding scenes |
| Below 85 | You can see what you gave away |
For a file that YouTube or Instagram will compress again, aim for the low 90s or higher, so the platform’s encode has a clean starting point.
What VMAF doesn’t tell you
- It needs the original. VMAF is a full-reference metric: no source file, no score.
- It assumes a viewing setup. The default model is tuned for a 1080p screen at a normal viewing distance. Netflix also publishes separate models for phones and for 4K.
- It can be flattered. Sharpening can raise the score without making the picture better, which is why there’s also a “NEG” variant that discounts that kind of enhancement.
- It’s an average. A clip can score well overall and still have a few rough seconds.
It isn’t perfect, and no metric is. But it’s repeatable and much better than “trust me”.
VMAF vs PSNR and SSIM
PSNR measures raw pixel error, and SSIM measures structural similarity. Both are quick and useful, but neither was built to predict what people notice. In Netflix’s own published testing, VMAF tracked viewer ratings more closely across different kinds of content.
How Becheloon uses VMAF
- It encodes your file for the destination.
- It measures the result against your original with VMAF.
- If the score is under the threshold, the file is not delivered. The app tries a higher-quality setting instead, or tells you why it can’t reach the target.
That’s why you’ll see something like “Quality 96/100 · 87% smaller” next to each output, and why you’ll never see a number nobody measured.
Why not just pick a high bitrate?
Because the same bitrate means very different things for a static talking head and for confetti in the wind. Measuring the result is the only way to make a file as small as it can be and still know it looks right. The guide to compressing a video without losing quality shows where the score fits in the whole process, and the 4 GB problem explains why phone footage has so much room to shrink. For VMAF used as a yardstick between encoders, see NVENC vs Quick Sync vs software encoding; for choosing between 4K and 1080p on YouTube, see how to compress 4K video for YouTube.
Want to see the number for one of your own clips? Becheloon runs on your computer, the report is free, and the first three videos get every feature.