Research-0126: `vmaf-tune` HDR dispatch coverage¶

Status: Active
Workstream: ADR-0300
Last updated: 2026-05-15

Question¶

vmaftune.hdr.hdr_codec_args() was the central encode-side HDR contract, but it only dispatched libx264, libx265, libsvtav1, hevc_nvenc, and libvvenc. The codec registry now contains more HDR-capable encoder rows: AV1 NVENC, HEVC/AV1 QSV, HEVC/AV1 AMF, HEVC VideoToolbox, and software AV1 via libaom. The open question was whether to leave those rows unflagged until HDR model training lands, or widen encode-side signaling now.

Sources¶

ADR-0300 defines the single dispatch-table contract and explicitly separates encode-side HDR correctness from the missing HDR VMAF model.
Research-0072 records the original HDR flag families: global -color_*, x265 private params, SVT-AV1 params, and HEVC NVENC 10-bit output.
Research-0087 inventories the adapter registry shape that hdr_codec_args() needs to cover.

Findings¶

Leaving the newer adapters out of hdr_codec_args() made the auto planner and corpus path record hdr_* provenance while emitting an empty hdr_args tuple for valid HDR-capable hardware codecs. That is an encode-side correctness gap independent of CHUG or Netflix's future HDR model release.

The least risky widening is:

use codec-private HDR SEI flags only where the fork already has a stable mapping (libx265, libsvtav1, hevc_nvenc);
force 10-bit output for hardware HEVC (p010le + main10);
force 10-bit 4:2:0 output for hardware AV1 (p010le);
carry FFmpeg-global -color_* tags for software AV1 and VVenC, where the in-tree evidence does not pin a stable private SEI option family.

Alternatives explored¶

Option	Verdict	Reason
Keep the narrow table until CHUG training completes	Rejected	CHUG blocks model training, not encode-side color signaling. Empty HDR args for hardware HDR rows are avoidable now.
Move HDR argv into each codec adapter	Rejected	ADR-0300 deliberately made `hdr_codec_args()` the single auditable contract; scattering the logic would reintroduce per-adapter drift.
Guess private SEI flags for every hardware backend	Rejected	Some FFmpeg wrappers expose different or unstable private option names. Global color tags plus 10-bit output are the safe floor.

Open questions¶

Add backend-specific mastering-display / MaxCLL private flags for QSV, AMF, VideoToolbox, and AV1 NVENC only after a real FFmpeg build matrix verifies those option names.
HDR VMAF scoring remains blocked until either Netflix releases a licensed model artifact or the CHUG pipeline produces a fork-owned HDR MOS head suitable for a separate model PR.

Research-0126: vmaf-tune HDR dispatch coverage¶