shreshthsaini/brightrate-study-gemma4-12b-native-pq

Gemma-4-12B BrightVQ native PQ study adapter

This PEFT adapter is one result from the BrightRate-LM controlled input and scaling study.

Base model

google/gemma-4-12B-it

Input interface

Eight 16-bit BT.2020 RGB frames remain in PQ code space. Raw PQ patches enter the learned patch projection while the pre-projection LayerNorm is bypassed.

Training data and recipe

This is the native PQ negative-result adapter from split 0. Training uses two epochs, a three-epoch cosine schedule horizon, learning rate 1e-4, micro-batch 1, gradient accumulation 8, and rank-16 LoRA with alpha 32 and dropout 0.05. MOS targets are interpolated across five quality words. The patch projection and vision-to-language projection are also trainable.

Training data: BrightVQ.

Metrics

On the 420-video split-0 test set: SROCC 0.0100, PLCC 0.0251, KRCC 0.0047, RMSE 13.3218.

Intended use

This adapter is provided to reproduce the native PQ negative result. It is not a recommended quality predictor. The matching code path is in src/native_pq_input.py.

Code and input construction are available in BrightRate-LM.

Citation

@article{saini2026brightratelm,
  title   = {BrightRate-LM: Representation-Aware Quality Assessment for User-Generated HDR Video},
  author  = {Saini, Shreshth and Wang, Yilin and Birkbeck, Neil and Adsumilli, Balu and Bovik, Alan C.},
  journal = {Machine Vision and Applications},
  year    = {2026},
  note    = {Submitted}
}