$ show gdavila/easyVmaf
The repository provides a command-line tool for computing VMAF (Video Multi-method Assessment Fusion) scores between reference and distorted video streams. It automates preprocessing steps required for accurate VMAF assessment, including deinterlacing, scaling, frame rate adaptation, and temporal synchronization. The tool supports both CPU and GPU-accelerated VMAF computation via libvmaf and libvmaf_cuda, and can process multiple video pairs using glob patterns. The architecture follows a three-layer design: the CLI layer handles argument parsing and output formatting, the VMAF layer orchestrates preprocessing and scoring, and the FFmpeg layer provides low-level wrappers for FFmpeg/FFprobe subprocesses and filterchain management. This separation of concerns makes the FFmpeg wrapper component particularly reusable for other multimedia processing tasks. The tool outputs results in JSON, XML, CSV, or to stdout, and includes features such as automatic sync window search, reverse sync, end-sync option, VMAF model selection (HD and 4K), and CAMBI heatmap generation. It is designed for Linux and macOS, with Docker support for both CPU and GPU environments.
The LLM identified 2 files worth highlighting for reuse:
This module provides a clean abstraction over FFmpeg and FFprobe subprocesses, managing probe operations and complex filterchain construction for video processing. It includes classes for probing video properties (FFprobe), building filtergraphs for PSNR/VMAF computation (FFmpegQos), and managing per-input filter chains (scaling, deinterlacing, frame rate conversion, trimming, and hardware upload). The component is isolated, depending only on configuration and standard library, and demonstrates non-obvious skill in handling filterchain ordering and hardware acceleration dependencies.
This module orchestrates the end-to-end VMAF computation process, including determining necessary preprocessing steps (based on video properties), performing temporal synchronization via parallel PSNR computation at different offsets, and constructing the filterchains for VMAF scoring. It encapsulates the logic for handling video alignment and preprocessing, making it a reusable pattern for video quality assessment tasks that require synchronization. The component demonstrates skill in multimedia processing orchestration and efficient parallel processing for synchronization search.