Netflix VMAF: Perceptual Video Quality Assessment via Multi-Method Fusion
An open-source framework for evaluating video quality using advanced perceptual metrics and models.
1Core VMAF Framework Overview
VMAF is an Emmy-winning perceptual video quality assessment algorithm developed by Netflix to measure visual fidelity. The software package provides a stand-alone C library called libvmaf, alongside Python bindings. It enables developers to train custom models, ensuring that video compression quality aligns closely with human visual perception rather than just mathematical error.
2Multi-Metric Integration Capabilities
Beyond its primary VMAF algorithm, the library includes implementations of several industry-standard metrics. Users can access PSNR, PSNR-HVS, SSIM, MS-SSIM, and CIEDE2000 within the same package. This integration allows engineers to perform comprehensive quality analysis by comparing multiple metrics simultaneously, providing a robust baseline for evaluating various video encoding and processing pipelines.
3Evolution of Model Versions
Netflix recently introduced a new set of VMAF models, designated as v1, to replace the legacy v0 generation. These models are documented with specific selection guidance to help users choose the right version for their use cases. The team maintains an open feedback loop, continuously updating these models to improve accuracy over time.
4Performance and API Optimizations
The release of libvmaf v3.0.0 marked a significant milestone by removing deprecated APIs from v2.0.0 and introducing critical bug fixes. Previous iterations achieved a 2x speed improvement through fixed-point and x86 SIMD-optimized implementations, specifically utilizing AVX2 and AVX-512 instructions. These architectural changes ensure the library remains efficient for high-throughput production environments.
5Detecting Banding with CAMBI
Netflix open-sourced the Contrast Aware Multiscale Banding Index, known as CAMBI, to detect contouring artifacts in video streams. Recent updates added full-reference input parameters and increased performance, achieving roughly 4.5x speed gains for 4k resolution content. This tool is essential for identifying visual degradation that standard metrics might overlook during the encoding process.
6No Enhancement Gain Mode
The introduction of the No Enhancement Gain mode addresses specific challenges when image enhancement operations are applied to video. This mode ensures that codec evaluation remains accurate by neutralizing artificial gains that could skew quality scores. It reflects Netflix's ongoing research into how post-processing operations impact the reliability of perceptual quality metrics.
7Licensing and Contribution Model
The project transitioned its licensing from Apache 2.0 to the BSD+Patent license to provide a more permissive framework with an express patent grant. Developers are encouraged to contribute to the codebase through established guidelines. The repository includes detailed documentation, slide decks, and contribution guides to facilitate community involvement and collaborative software development.
8Strategic Importance for Streaming
VMAF serves as a critical tool for streaming services to balance bandwidth consumption with visual quality. By leveraging multi-method fusion, it provides a nuanced understanding of how compression affects the viewer experience. This framework remains a standard for codec evaluation, enabling data-driven decisions that directly impact the quality of service for global users.