BrightRate-LM

Representation-Aware Quality Assessment for User-Generated HDR Video

Shreshth Saini, Yilin Wang, Neil Birkbeck, Balu Adsumilli, Alan C. Bovik

What this is

BrightRate-LM predicts the perceptual quality of user-generated HDR video and returns a score, a description of visible defects, and reasoning for the score. This work extends BrightRate (WACV 2026) by testing how multimodal language models respond to tone-mapped, multi-exposure, and native PQ inputs.

Key figures

Accuracy, memory, and throughput comparisons for selected models — BrightRate-LM gives the strongest measured accuracy among the selected models while retaining a 7B backbone.

SROCC change from SDR to multi-exposure input — Multi-exposure input improves four of five matched model and recipe pairs.

Zero-shot and supervised scaling across model families — Supervised quality prediction does not improve monotonically with model size.

Training loss curves for the studied model families — The matched recipes converge, but lower training loss does not guarantee better held-out ranking.

BrightRate-LM score, description, and reasoning example — The output interface pairs the numeric estimate with a short visual diagnosis.

Main result

BrightRate-LM values are five-split means. BrightRate values are published 100-split medians.

Model	Input	SROCC	PLCC	KRCC	RMSE
BrightRate-LM, 7B	Multi-exposure	0.9052	0.9107	0.7281	5.5348
BrightRate, published	HDR-aware features	0.8887	0.8970	0.7059	5.7514

Insights

The multi-exposure interface improves four of five matched models.
The largest measured gain is 0.0638 SROCC for Qwen2.5-VL-3B.
Qwen3-VL-2B with multi-exposure input exceeds Qwen3-VL-8B with tone-mapped input.
Gemma-4 zero-shot strength does not carry over to the matched supervised recipe.
Both tested native PQ interfaces fail to learn a useful ranking.
BrightRate-LM reports a score, a defect description, and score reasoning.

Data

Experiments use BrightVQ from BrightRate.

BibTeX

@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}
}