Netflix/vmaf

Perceptual video quality assessment based on multi-method fusion.

What it solves

It provides a way to objectively measure the perceptual quality of video, helping developers and engineers determine how much a video's quality has degraded due to compression or other processing without relying solely on human observation.

How it works

VMAF (Video Multi-Method Assessment Fusion) is a perceptual video quality assessment algorithm. It combines multiple metrics into a single score. The project provides a C library (libvmaf) for high-performance implementation, a Python library for training and testing custom models, and a command-line tool for production deployment. It also includes other standard metrics like PSNR, SSIM, and MS-SSIM, as well as CAMBI for detecting banding artifacts.

Who it’s for

It is designed for video engineers, codec developers, and quality assurance teams who need to measure video quality accurately and efficiently, particularly those integrating these metrics into FFmpeg or production pipelines.

Highlights

  • Emmy-winning perceptual quality algorithm developed by Netflix.
  • Includes a specialized detector for banding artifacts called CAMBI.
  • Integrated as a filter in FFmpeg.
  • Offers a No Enhancement Gain (NEG) mode to ensure codec evaluations are not skewed by image enhancement operations.
  • High-performance implementation with x86 SIMD optimizations (AVX2, AVX-512).

Related

  • Project
    NVIDIA-AI-Blueprints/video-search-and-summarization

    NVIDIA’s Video Search & Summarization (VSS) blueprint is a GPU‑accelerated reference stack that lets you search, query, and summarise live or recorded video with natural‑language. It combines vision‑language models, LLMs, and NVIDIA NIM micro‑services, offering real‑time analytics, alert verification, and long‑video summarisation via modular Python/Node/Next.js services and Docker‑Compose or cloud‑launchable deployment.

  • Project
    Linfeng-Tang/SeAFusion

    A semantic-aware real-time network that fuses infrared and visible images to optimize them for high-level vision tasks like object detection and semantic segmentation.

  • Project
    LAION-AI/CLIP_benchmark

    A standardized evaluation framework for CLIP-like models to measure performance on zero-shot classification, retrieval, captioning, and linear probing across diverse datasets.

  • Project
    ModelTC/LightX2V

    LightX2V is an open‑source inference framework for fast image and video generation (T2V, I2V, T2I, I2I, T2AV, etc.). It supports many state‑of‑the‑art models (MiniMax‑H3, Wan, HunyuanVideo, Qwen‑Image, etc.) and provides speed‑up techniques such as 4‑step distilled LoRAs, FP8/NVFP4 quantization, tensor/sequence parallelism, and disaggregated deployment on a wide range of hardware. Benchmarks show up to 3.9× faster inference than competing frameworks on H100 GPUs. The project offers Docker images, pip install, Gradio/ComfyUI front‑ends, extensive docs, and an online demo.

  • Project
    tschnz/Live-Video-Magnification

    A real-time video magnification tool that amplifies subtle motion and color changes using Eulerian video magnification techniques.