Skip to content

Research-0653: CHUG Display Profile Training

Question

How should CHUG HDR MOS training consume display and panel context without turning a local viewing setup into corpus source data?

Findings

  • CHUG gives the fork HDR subjective MOS rows, but the current feature rows are still clip and ladder facts. They do not describe the target display used for local model experiments.
  • The signal-mix audit identified HDR display and panel context as a blind spot. The missing signals are not more canonical-6 statistics; they are viewing-context variables such as peak luminance, black level, color-volume coverage, ambient light, panel class, local dimming, and tone-mapping.
  • Hard-coding one display into CHUG JSONL would make the corpus look more authoritative than it is and would require row rewrites for every target panel experiment.
  • A trainer-side profile keeps the source rows unchanged while making the experiment reproducible through the emitted manifest.
  • Future multi-display HDR corpora need row-local display values to win over a global fallback profile; otherwise those datasets would lose their main distinguishing axis.

Result

Use a --display-profile-json training input and a named chug-hdr-display-v1 schema. The profile normalizer accepts common raw field names (peak_nits, black_nits, ambient_lux, panel_type, tone_mapping, BT.2020/P3 coverage) and projects them onto stable numeric feature columns. The CHUG wrapper auto-selects the display-aware schema only when the profile flag is supplied and no explicit schema was requested.

Verification

Regression tests cover profile normalization, row-local display value precedence, the chug-hdr-display-v1 feature order, CHUG wrapper schema selection, and manifest recording of the normalized profile plus source sha256.