Modernized perceptual video quality assessment.
A cloud-native fork-descended-from Netflix VMAF, rebuilt for distributed video pipelines: GPU-vendor-agnostic backends, tiny-AI models for one-shot scoring, MCP server for AI-agent integrations, Kubernetes-native deployment with a 3-vendor GPU device-plugin matrix.
- vmafx — the main repo: core C++23 library (
libvmafC ABI being retired), Go production tooling, Rust pilots, Python ML training, Helm chart, dev container.
- Cloud-native first. Production artifacts are Docker images + Helm chart. Native
.so/.deb/.rpmpublishing is out of scope — Docker is the platform. - Multi-vendor GPU. First-class support for NVIDIA (CUDA), AMD (HIP / ROCm), Intel (SYCL / Arc).
- Correctness above speed. Bit-exactness against the Netflix CPU reference is the non-negotiable safety net.
- Multi-language, multi-purpose. Go for production tooling (controller / node / Operator / CLI / MCP), Rust for pilot integrations + bindings, Python for ML training, C++23 internals for the scoring core.
Dual-licensed for fork-added code: BSD-3-Clause-Plus-Patent OR MIT. Netflix-inherited code stays under BSD-3-Clause-Plus-Patent. See the LICENSE and LICENSE-MIT on the main repo.
Active development. The fork is undergoing an aggressive modernization sweep (Phase 4) toward a distributed controller / node / operator architecture. See ADR-0686 (rebrand umbrella) and ADR-0709 (Phase 4b distributed platform).