updated
Sep 16, 2026
4.27MB
file size
4.9K
downloads
VMAF sits between encoding and final output, giving everyday users an objective way to check whether their video quality survived the compression process.
Netflix VMAF scores how much quality a compressed video actually lost, on a scale of 0 to 100, by comparing it against the original.
It is the same scorer Netflix built for its own encoding pipeline, and it is free and open source.
You can download Netflix VMAF here, along with a third-party graphical front end if you would rather not use a command line.
What the Score Actually Means
VMAF does not compare the two files as raw data - it predicts what a human viewer would say about the difference. These are the bands people work to in practice.
| Score | What you are looking at | What to do |
|---|---|---|
| 95 and above | No difference you could find in normal viewing | Stop here - going higher only costs file size |
| 90 to 95 | Effectively transparent for most people, most of the time | The usual target for a copy you intend to keep |
| 80 to 90 | Good. Softness and lost fine detail show up if you pause and look | Fine for casual viewing, re-encode if it is replacing the original |
| 70 to 80 | Noticeably softer, with blocking in dark or busy scenes | Raise the quality setting one step and score it again |
| Below 70 | Clearly degraded | Re-encode - the compression went too far |
One thing to know before you trust a number: the default model, vmaf_v0.6.1, was trained on 1080p video watched on a television. Comparing phone-sized or 4K video means switching to the matching model, or the bands above are being used outside what they were built for.
It Will Not Open Your MP4
This is the part that catches everyone. The VMAF program itself does not read MP4, MKV or any other normal video file - it accepts only .y4m and raw .yuv, and raw YUV additionally needs you to state the width, height, pixel format and bit depth by hand.
So something has to decode your files first. That is what the two easier routes below do for you.
Both files also have to match in resolution and frame count. If you encoded at a lower resolution, it has to be scaled back up to the original's size before the comparison means anything, and MediaInfo is the quickest way to confirm the two files line up.
The Three Ways to Run It
- FFmpeg, with the libvmaf filter. The easiest route by a distance - FFmpeg decodes both files itself, so it works on the MP4s you already have, in one command.
- The VMAF GUI on the download page. Version 2.1.4, a free MIT-licensed C# front end by ThatNerdUKnow - but you supply FFmpeg, the VMAF build and the model files yourself, and it writes YUV files to disk as it works.
- The command-line tool on its own. Correct and fastest, but you are converting to Y4M first and typing every parameter, which is only worth it if you are scripting a batch.
One thing to expect from the GUI: it shows no progress while it is scoring, so on a long clip it looks like it has frozen when it has not.
If the choice between a window and a command line is the thing putting you off, our guide on GUI versus command line is a fair summary of what you gain either way.
Where It Fits in Your Encoding Workflow
VMAF replaces nothing. It is the check you run after the encode, and its real value is on a test clip rather than a finished file.
Pick a twenty-second segment with motion and some dark detail, encode it two or three times at different quality settings, score each one, and keep the cheapest setting that clears your threshold. Then run your whole library at that setting instead of guessing and finding out later.
That works with whatever you already use, and for most people that is HandBrake.
For batch work, FastFlix and MediaCoder both let you queue a set of test encodes in one go.
For encoder-level control, x265 HEVC Encoder and Av1ador are the ones worth scoring carefully, because HEVC and AV1 are exactly where quality loss is subtle enough that eyeballing it fails.
What VMAF Will Not Tell You
A score is a prediction of average human opinion, not a verdict, and there are four things it quietly leaves out.
- Audio. VMAF scores picture only, so an encode that scores 96 can still have mangled the soundtrack.
- Filters that flatter the metric. Sharpening can push a score up while making the result look worse to an actual person - a high number earned that way is not a better encode.
- One number for a whole clip. An average of 92 can hide one badly broken scene, which is why short test segments beat scoring a full film.
- Anything about playback. A file can score well and still stutter on your hardware.
So watch the result as well as scoring it - MPC-BE with LAV Filters gives accurate playback for exactly that.
And our Codec Checker reads what is actually inside a finished file, straight from the browser.
Quick questions
Is Netflix VMAF free?
Yes. It is open source from Netflix, and the graphical front end offered alongside it is free and MIT licensed as well.
Why will it not open my MP4?
Because the program itself only reads Y4M and raw YUV. Use FFmpeg's libvmaf filter, or the GUI, and let those decode your files for you.
What score should I aim for?
90 to 95 is the usual target for a copy you plan to keep. Above 95 you are mostly paying in file size for a difference you will not see.
Do the two files need to be the same resolution?
Yes, and the same frame count. A lower-resolution encode has to be scaled back up to the original's size first, or the comparison is not measuring what you think it is.
Can I trust the score on its own?
Use it as a guide, not a verdict. It ignores audio, it can be nudged upward by sharpening, and an average across a whole clip can hide one bad scene.
Worth reading next: StaxRip against HandBrake if you are still choosing an encoder, and the best free video codecs for HD streaming for the settings side of it.
Download Netflix VMAF 3.2.1 and score your next encode before you commit to it, or tell us what score you got if you have used it before.
