VMAF
VMAF (Video Multimethod Assessment Fusion) is a video quality metric, developed by Netflix with university researchers, that predicts how good a compressed video looks to a human viewer. VMAF compares the encoded video with its original source and gives a score from 0 to 100. Higher scores mean the encoded version looks closer to the original.
Why it matters
VMAF lets encoding teams cut bitrate, and with it storage and delivery bills, while checking that viewers will not notice a drop in quality. It replaces slow human viewing tests with an automated score that runs across whole libraries. That makes quality versus cost decisions measurable.
How it works
- The original source video and the compressed version are compared frame by frame.
- Several basic measurements check detail loss, visual information fidelity and motion between frames.
- A machine learning model, trained on human quality ratings, fuses those measurements into one score per frame.
- Frame scores are pooled into a single 0 to 100 score for the whole video.
What is a good VMAF score?
There is no official pass mark, but higher is better and a score in the 90s usually means few viewers would notice a difference from the source. Teams typically set their own targets per rendition, with the highest quality level aiming highest and phone-sized renditions allowed to sit lower. Scores depend on the model used, such as the default or phone model, so compare like with like.
VMAF vs PSNR vs SSIM: which should I use?
Use VMAF when you want a score that tracks what viewers see, and keep PSNR and SSIM as fast supporting checks. PSNR measures raw pixel error, SSIM measures structural similarity and VMAF combines several measures with a model trained on human ratings. Many encoding teams report all three, because each catches problems the others miss.
Can VMAF be used for live video?
VMAF can be used for live video, but it needs the original source to compare against, so it usually runs on test streams or sampled feeds rather than every viewer's stream. It is also heavier to compute than PSNR. Live teams often use it to tune encoder settings before events, then monitor simpler signals during the broadcast.
Where it's used
Encoding teams at OTT services, live sports streamers and video platforms use VMAF to tune bitrate ladders and per-title encoding, shrinking files and bandwidth bills without a visible drop in quality.
By the numbers
- Netflix announced VMAF on its Tech Blog in June 2016, developed with the University of Southern California, Nantes University and the University of Texas at Austin. (Wikipedia: Video Multimethod Assessment Fusion)
- VMAF scores each frame from 0 to 100, where 100 means quality identical to the reference video. (Wikipedia: Video Multimethod Assessment Fusion)
- VMAF was recognized with a Technology and Engineering Emmy Award in 2021, shared by Netflix and its university partners. (Wikipedia: Video Multimethod Assessment Fusion)
