vmaf

# VMAF - Video Multi-Method Assessment Fusion

[![libvmaf](https://github.com/Netflix/vmaf/actions/workflows/libvmaf.yml/badge.svg)](https://github.com/Netflix/vmaf/actions/workflows/libvmaf.yml)
[![Windows](https://github.com/Netflix/vmaf/actions/workflows/windows.yml/badge.svg)](https://github.com/Netflix/vmaf/actions/workflows/windows.yml)
[![ffmpeg](https://github.com/Netflix/vmaf/actions/workflows/ffmpeg.yml/badge.svg)](https://github.com/Netflix/vmaf/actions/workflows/ffmpeg.yml)
[![Docker](https://github.com/Netflix/vmaf/actions/workflows/docker.yml/badge.svg)](https://github.com/Netflix/vmaf/actions/workflows/docker.yml)

VMAF is an [Emmy-winning](https://theemmys.tv/) perceptual video quality assessment algorithm developed by Netflix. This software package includes a stand-alone C library `libvmaf` and its wrapping Python library. The Python library also provides a set of tools that allows a user to train and test a custom VMAF model.

Read [this](https://medium.com/netflix-techblog/toward-a-practical-perceptual-video-quality-metric-653f208b9652) tech blog post for an overview, [this](https://medium.com/netflix-techblog/vmaf-the-journey-continues-44b51ee9ed12) post for the tips of best practices, and [this](https://netflixtechblog.com/toward-a-better-quality-metric-for-the-video-community-7ed94e752a30) post for our latest efforts on speed optimization, new API design and the introduction of a codec evaluation-friendly [NEG mode](resource/doc/models_v0.md#disabling-enhancement-gain-neg-mode).

Also included in `libvmaf` are implementations of several other metrics: PSNR, PSNR-HVS, SSIM, MS-SSIM and CIEDE2000.

![vmaf logo](resource/images/vmaf_logo.jpg)

## News

- (2026-10-26) We are releasing an HDR version for VMAF-v1. Please refer to [models_v1.md](resource/doc/models_v1.md#hdr-models) for more information.
- (2026-06) We are releasing a new set of VMAF models (**v1**). See [tech blog](https://medium.com/netflix-techblog/vmaf-v1-good-is-not-good-enough-60d7e4244ea8) for more information on v1. See [models_v1.md](resource/doc/models_v1.md) for details and model selection guidance. The previous generation of models (v0) remains documented in [models_v0.md](resource/doc/models_v0.md). We are open to user feedback on VMAF v1 and will continue to improve and update the models over time.
- (2023-12-07) We are releasing `libvmaf v3.0.0`. It contains several optimizations and bug fixes, and a full removal of the APIs which were deprecated in `v2.0.0`.
- (2021-12-15) We have added to CAMBI the `full_ref` input parameter to allow running CAMBI as a full-reference metric, taking into account the banding that was already present on the source. Check out the [usage](resource/doc/cambi.md) page.
- (2021-12-1) We have added to CAMBI the `max_log_contrast` input parameter to allow to capture banding with higher contrasts than the default. We have also sped up CAMBI (e.g., around 4.5x for 4k). Check out the [usage](resource/doc/cambi.md) page.
- (2021