What is VMAF and How to Use It for Video Quality

What is VMAF?

VMAF — Video Multi-Method Assessment Fusion — is a video quality metric developed by Netflix to predict how good a video looks to actual human viewers. Instead of relying on a single formula, it fuses several quality measures together using a model trained on real human opinion scores. The result is a number that tracks perceived quality far better than older metrics on their own.

How to read a VMAF score

VMAF is reported on a 0–100 scale, where higher is better:

• 95–100 — visually indistinguishable from the source for most viewers.
• 80–95 — good quality; minor artefacts only noticeable on close inspection.
• 60–80 — acceptable but visibly degraded.
• Below 60 — clearly degraded quality.

Roughly, every ~6 VMAF points corresponds to a noticeable step in perceived quality. It’s a “full-reference” metric, meaning it compares an encoded video against the original source.

VMAF vs PSNR vs SSIM

• PSNR measures raw pixel error. Simple and fast, but it doesn’t match human perception well — two videos with the same PSNR can look very different.
• SSIM measures structural similarity (contrast, luminance, structure). Much closer to perception than PSNR.
• VMAF goes furthest: it’s trained on human opinion scores and fuses multiple measures, so it’s the best single predictor of perceived quality — which is why it’s the industry standard for encoding decisions.

Use PSNR/SSIM for quick checks; use VMAF when you care about how the video actually looks to viewers.

Getting a quality score without running VMAF yourself

VMAF’s own reference implementation (libvmaf) is accurate but heavy — computing it is one of the more CPU-intensive things you can do to a video, which is why it usually runs on a dedicated encoding pipeline rather than as a quick web check. DiffALL’s video comparison tool doesn’t run VMAF; it scores every comparison with per-second SSIM and PSNR instead, plus a difference heatmap that shows exactly where two videos diverge:

1. Upload the original and the encoded/compressed version.
2. DiffALL analyses the frames and reports an overall SSIM score, a per-second PSNR/SSIM chart, and a heatmap of the worst-differing frame.
3. Use the score and the heatmap together to decide whether your encoding settings preserved enough quality — and to see exactly where they didn’t.

SSIM and PSNR won’t match a VMAF score number-for-number — they’re a different (if related) family of metrics — but for the everyday question of “did this re-encode hold up, and where did it fall apart,” they get you an answer in seconds instead of a dedicated encoding run.

Common use cases

• Encoding decisions — compare bitrate or codec settings and pick the one that keeps SSIM/PSNR high.
• Transcoding QA — confirm a converted video didn’t lose perceptible quality, and see exactly which frames did.
• Streaming optimisation — balance file size against measured quality without a dedicated VMAF pipeline.

Try it now

Want a quality score and a heatmap for your own footage? Upload the original and the encoded version and let DiffALL measure exactly how much quality you kept — and show you where.