Netflix's Emmy-winning VMAF ships new models and keeps custom training tools open

The story

Netflix's VMAF, the Emmy-winning open-source library for judging video quality the way human viewers do, has surfaced on Hacker News's front page. The project's GitHub page shows a codebase that is still very much alive: a new generation of quality models arrived in June 2026, and the package continues to offer something unusual in this space — the ability for anyone to train and test a custom quality model on their own footage.

What VMAF does

VMAF, generally expanded as Video Multi-method Assessment Fusion, is Netflix's algorithm for perceptual video quality assessment built on multi-method fusion. Instead of trusting a single measurement, it combines several of them into a score intended to track how people actually perceive picture quality. The work earned the project an Emmy award, according to the repository.

The package has several layers. libvmaf is a standalone C library for embedding the metric in other software, and a Python library wraps it with scripts for testing, model training and validation, dataset processing and data visualisation. A command-line tool provides the complete implementation for production deployments. VMAF is also integrated into FFmpeg as a filter, enabled at configure time, and the project ships a Docker image plus Windows build instructions.

Alongside the fused metric itself, libvmaf includes implementations of PSNR, PSNR-HVS, SSIM, MS-SSIM and CIEDE2000, so teams can benchmark the learned score against the classical alternatives.

New models and recent history

According to the project's news section, Netflix released a new set of models, labelled v1, in June 2026, with documentation offering model-selection guidance. The previous generation, v0, remains documented, and Netflix says it will keep refining v1 based on user feedback.

The library has evolved steadily underneath. libvmaf v3.0.0 arrived in December 2023 with optimisations, bug fixes and the removal of APIs deprecated back in the v2.0.0 release. That v2.0.0 version, from December 2020, introduced a fixed-point implementation with x86 SIMD optimisations — AVX2 and AVX-512 — that roughly doubled throughput compared with the earlier floating-point code, along with a new, more extensible API.

Two other additions stand out. A No Enhancement Gain (NEG) mode was added in July 2020 to make the metric fairer when evaluating codecs, addressing how the score behaves in the presence of image enhancement. And in October 2021 Netflix open-sourced CAMBI, its Contrast Aware Multiscale Banding Index, a detector for the banding or contouring artefacts that show up in smooth gradients. CAMBI subsequently gained parameters for capturing higher-contrast banding and for full-reference operation, plus a speed-up of around 4.5x at 4K resolution, per the project's changelog. The licence also moved from Apache 2.0 to BSD+Patent in February 2020, a switch Netflix characterises as more permissive and inclusive of an express patent grant.

Train your own quality model

The feature that keeps VMAF relevant beyond Netflix is the Python tooling for training and testing custom models. Perceived quality varies with content type — animation, live sport and film grain stress encoders in very different ways — so a metric tuned on one catalogue may not be optimal for another. With dataset-processing, training and validation scripts bundled in the repository, other streaming services and encoder vendors can fit the model to their own subjective test data rather than relying on Netflix's defaults.

Adoption in codec testing

VMAF has also spread into standards work. The Alliance for Open Media has specified it as the standard implementation metrics tool under the AOM common test conditions, the framework used to compare codec implementations against one another in a repeatable way.

Why it matters

Streaming video is a constant trade-off between bitrate and perceived quality, and a metric that misjudges human perception produces bad encoding decisions — either wasted bandwidth or visible degradation. PSNR, the traditional fallback, is widely known to correlate poorly with viewer opinion. An open, Emmy-recognised metric that can be run through FFmpeg, embedded via a C library, and retrained on custom content gives smaller platforms access to tooling that was once the preserve of companies with Netflix's research budget. The June 2026 v1 release, with Netflix actively soliciting feedback, signals a maintained project rather than an archived one — and its role in AOM's test conditions means improvements here effectively shape how future video codecs are judged.