VMAF

VMAF, or Video Multimethod Assessment Fusion, is a full-reference perceptual video quality metric designed to predict human judgments of video quality. It aims to provide a single score that correlates with subjective opinions by fusing multiple quality indicators through a regression model trained on MOS (mean opinion score) data.

The metric combines outputs from several base quality measurements—such as Visual Information Fidelity (VIF), structural similarity measures, and other perceptual features—along with temporal information to capture motion and flicker. A machine-learning fusion model integrates these features into a final VMAF score, typically scaled from 0 to 100, where higher values indicate better quality.

VMAF was developed by Netflix with contributions from researchers at academic institutions and released as an open-source project in 2015. The reference implementation is available on GitHub and supports common computing platforms, with implementations in C and Python to run on Linux, Windows, and macOS. The score is designed to reflect perceived quality across a range of content types, bitrates, resolutions, and color depths.

Applications of VMAF include benchmarking and comparing video encoders, guiding bitrate ladders and encoding parameters in streaming pipelines, and monitoring quality across content libraries. It has been widely adopted in the industry for codec evaluation and quality control due to its emphasis on perceptual relevance and its ability to align with subjective viewing experience. Users should note that VMAF relies on MOS-based training and may require calibration for specific displays, color spaces, or content characteristics to maintain predictive accuracy.