VMAF, Video Multimethod Assessment Fusion, is the full-reference metric Netflix built and open-sourced to predict, on a 0-to-100 scale, how good a compressed video looks to a person. Instead of measuring one thing, it fuses several elementary features — in the classic v0 design Visual Information Fidelity (VIF), the detail-loss metric (ADM/DLM), and a motion feature — through a machine-learning model (a support vector regressor) trained on subjective opinion scores, so a jump from 60 to 70 means roughly as much as 80 to 90. The default model is vmaf_v0.6.1, with separate phone and 4K models for other viewing conditions. The catch is that the training that makes VMAF accurate is also its baggage: a bare score is meaningless without naming the model, the pooling, and the content, and it can be inflated by sharpening unless you use VMAF-NEG. It is full-reference, weakest off its training distribution, and the industry default for streaming quality.
The handbook · 7 volumes
Fora Soft Video Engineering Handbook
VMAF is the industry default for judging encodes, yet the model, device setting and pooling all change what the number means. Our seven-volume handbook covers the full video stack; for VMAF start with Volume 1, which explains VMAF and its models, quality targets, encoding ladders and comparing codecs, then Volume 3 for delivery. Written by Nikolay Sapunov, CEO at Fora Soft.
Using VMAF to set encoding ladders and not sure the scores are comparable?
We will audit your VMAF setup, models and pooling and give you a scoring method you can trust for ladder decisions. 250+ video products shipped since 2005.
