- Published on
Netflix vmaf: Perceptual video quality assessment based on multi-method fusion
- Authors
- Name
- DeepHorus
Netflix has released a new generation of VMAF models, version 1 (v1), as of June 2026. VMAF is an Emmy-winning perceptual video quality assessment algorithm. The new models aim to enhance video quality evaluation for streaming and codec development.
VMAF was developed by Netflix and includes a stand-alone C library, libvmaf, and a Python wrapper. The Python library allows users to train and test custom VMAF models. It also implements other metrics: PSNR, PSNR-HVS, SSIM, MS-SSIM, and CIEDE2000.
The v1 models are detailed in models_v1.md, while the previous v0 models remain in models_v0.md. In December 2023, Netflix released libvmaf v3.0.0, which removed deprecated APIs and included optimizations. Earlier, v2.0.0 (December 2020) achieved a 2x speedup with fixed-point and SIMD optimizations.
VMAF is integrated as an FFmpeg filter and is the standard implementation metric for AOM CTC. The command-line tool vmaf provides a complete algorithm implementation for production environments. This makes it crucial for codec evaluation and quality assurance. The new v1 models could lead to more accurate perceptual quality scores.
Netflix is open to user feedback on VMAF v1 and will continue to improve the models over time. The company also open-sourced CAMBI in October 2021 for banding detection. In July 2020, Netflix introduced the No Enhancement Gain (NEG) mode for codec evaluation.
Future developments will likely focus on further speed optimizations and model refinements. VMAF's license changed to BSD+Patent in February 2020, promoting broader adoption. As video streaming grows, VMAF remains a vital tool for ensuring high-quality experiences. Netflix continues to improve VMAF, with v1 models now available for testing.
