Documentation¶

This is an overview of the available documentation in the VMAF repository.

Getting started¶

• Installation – per-OS install scripts (Ubuntu, Fedora, Arch, Alpine, macOS, Windows)
• Building on Windows – build VMAF from source on Windows (upstream guide)

Usage¶

• CLI reference – the vmaf command-line tool, every flag, defaults, examples
• vmaf_bench – micro-benchmark & GPU-vs-CPU validation harness
• --precision – score output precision (default %.6f Netflix-compat; opt into %.17g round-trip lossless via --precision=max)
• Python library – explains the Python wrapper for VMAF
• FFmpeg – how to use VMAF in conjunction with FFmpeg
• Docker – how to run VMAF with Docker
• MATLAB – running other quality algorithms (ST-RRED, ST-MAD, SpEED-QA, and BRISQUE) with MATLAB
• BD-rate utilities – Bjontegaard-delta rate and VMAF gain helpers
• Per-shot scoring – per-shot VMAF scoring workflow
• ROI scoring – region-of-interest saliency-weighted scoring
• External resources – e.g. software using VMAF
• vmaf-tune — quality-aware encode automation harness: overview | fast path | bitrate ladder | codec adapters | recommend | saliency-aware | resolution-aware | HDR & sampling | cache | bisect

C API¶

• API overview – core libvmaf.h + picture.h + model.h + feature.h: context lifecycle, scoring, pictures, models, ABI-stability tiers, thread-safety, runnable example
• DNN sessions – tiny-AI dnn.h: standalone ONNX sessions (luma filter + multi-input named binding), device config, error codes
• GPU (CUDA / SYCL) – libvmaf_cuda.h + libvmaf_sycl.h: zero-copy frame buffers, dmabuf / VA / D3D11 import, profiling
• MCP C API – libvmaf_mcp.h: embedded in-process MCP server (stdio / UDS / SSE transports)

Metrics¶

• Features – VMAF's core features (metrics)
• CAMBI – contrast-aware multiscale banding index
• SSIMULACRA 2 – fork-added structural similarity metric (modern-codec quality)
• DISTS-Sq – fork-added deep image structure & texture similarity extractor (FR; smoke checkpoint)
• Confidence Interval – bootstrapping for CI estimates of VMAF scores
• Bad Cases – how to report cases of VMAF not working well
• AOM CTC – running VMAF under AOM common test conditions
• NFLX CTC – running VMAF under NFLX common test conditions

Models¶

• Overview – summary of the available pre-trained models
• Datasets – the two publicly available datasets for training custom models

Backends¶

GPU / SIMD backend notes under backends/:

Architecture¶

• Repository layout – what lives where + decision tree
• Python-harness workspace – the moved workspace/ tree
• ADR log – every non-trivial architectural / policy decision + rationale

MCP¶

• MCP server overview – install, security model (path allowlist), env vars, Claude Desktop / Cursor config
• Tool reference – per-tool request/response schemas + error codes for vmaf_score, list_models, list_backends, run_benchmark, eval_model_on_split, compare_models
• Embedded server – in-process libvmaf_mcp.h server (stdio / UDS / SSE); compute_vmaf + list_features tools live
• Release channel – PyPI packaging and versioning for the standalone Python server

Tiny-AI¶

• Tiny-AI docs – overview, training, inference, benchmarks, security
• Training + training data + MOS corpora
• Inference – ONNX Runtime EPs: CPU, CUDA, TensorRT, CoreML, OpenVINO (CPU / GPU / NPU)
• LOSO evaluation + predictor + conformal VQA
• Ensemble training kit + ensemble v2 runbook
• Quantization + quant epsilon
• Model registry – canonical registry of all shipped ONNX / JSON models
• Tiny-AI roadmap – Wave 1 scope expansion (LPIPS, saliency, per-shot CRF, vmaf_post, allowlist Loop/If, MCP VLM tool)

Development¶

• Engineering principles – NASA Power-of-10 + JPL + CERT + MISRA, golden gate, quality policy
• Benchmarks – fork-added benchmark numbers (GPU, SIMD, --precision)
• Build flags – every meson_options.txt option with defaults, effects, and flag interactions
• Release – how to perform a new release
• CI – CI pipeline overview + required status checks
• CI runners – self-hosted runner setup for GPU lanes
• Cross-backend gate – T6-8 GPU-parity gate semantics
• Fuzzing – OSS-Fuzz integration
• IDE setup – clangd / VS Code configuration for all backends
• FFmpeg patches refresh – rebasing the ffmpeg-patches/ series
• Upstream watchers – tracking Netflix/vmaf divergence

Reference¶

• FAQ
• References – a list of links and papers
• Papers and Presentations