v1.4.1

Fixed

• Vertical videos are scored with the right model, upright
  (#2, #3).
  The automatic model looked at the height only: a 1080×1920 short got the
  4K model (vmaf_v1.0.16_3d0h_2160, which adds 5 points and clips at 110,
  hence scores above 100), and every portrait source was stretched to a
  landscape evaluation size. The model is now chosen on the short side of
  the picture, and a portrait source is evaluated upright (1080×1920). On
  the short of #2, the exact score matches Netflix's vmaf tool at the
  native size to the third decimal (97.113, 95.512 with vmaf_v0.6.1),
  instead of 100.7. Scores of portrait videos change; landscape ones do
  not. Ladders of portrait sources are measured the same way. Library:
  vmaf.ResolveModelForSize; vmaf.ResolveModel is deprecated.
• A container without a memory limit no longer gets killed for lack of
  memory (#1). Without --memory
  nor a cgroup limit (Docker Desktop's default), qc now sizes its workers
  on the memory the system has available (MemAvailable, Linux) instead
  of not at all. With the fix above, qc vmaf on a 1080×1920 short peaks
  at 1.5 GB with 6 workers, where 1.4.0 took more than 4 GB with 2.
• Docs: the memory of a vertical video, what the default --metrics cost,
  and why libvmaf 3.2.1 reports itself as 3.2.0.

Docker

docker run --rm -v "$PWD:/data" ghcr.io/eko/qc:1.4.1 version --check

Binaries

Self-contained (libvmaf and the VMAF models are built in); ffmpeg is still
needed. Linux binaries are static; the macOS one is not signed
(install).