vmaf-cuda

GPU-accelerated VMAF using CUDA and NVDEC, up to 4.4x throughput and 37x lower latency at 4K versus CPU, now folded into VMAF 3.0/FFmpeg 6.1.

Link

Related resources

  • Vship (Vapoursynth-HIP-SSIMU2)GPU-accelerated (HIP/CUDA) visual fidelity metrics library implementing SSIMULACRA2, Butteraugli, and CVVDP.
  • ffmpeg-webguiContainerized web GUI for FFmpeg with transcode job queue and built-in VMAF quality scoring.
  • ffmpeg-queueFastify + Redis job queue for dispatching FFmpeg transcoding jobs, Docker Compose ready, production-used.
  • av1-scdScene-change detection CLI generating encoder force-keyframe configs, wrapping selectable backends (pyscene, vs-xvid, av-scenechange, ffmpe…
  • ShotcutQt/MLT+FFmpeg-based cross-platform non-linear video editor, actively maintained with ~14k stars.
  • PXNR No-Reference Video Quality EstimationEricsson's sample implementation of the ITU-T P.NATS phase 2 no-reference video quality estimation module.
  • fssimu2Zig implementation of SSIMULACRA2 metric, ~23% faster and ~40% less memory than the reference implementation.
  • BVQA_BenchmarkBlind (no-reference) VQA benchmarking toolkit comparing multiple models.
  • CompressedVQA-AEVFull-reference + no-reference VQA for asymmetric-encoded video, QoMEX 2026 challenge solution using SigLIP2+Swin-B.
  • CAMP-VQAWACV 2026 caption-embedded multimodal no-reference VQA model for compressed video, with code and HF demo.