
Multimodal LLMs for HDR Video Quality Assessment
with HDR-Aware Policy Optimization
Shreshth Saini
Laboratory for Image & Video Engineering (LIVE) · The University of Texas at Austin
Advised by Prof. Alan C. Bovik
CVPR 2026 · December 2025
Saini et al., Seeing Beyond 8bits, CVPR 2026 (arXiv:2603.00938)
Right now the answer is no. Every existing model was built for the 8-bit SDR world.
Every modern phone captures 10-bit HDR by default.
HDR keeps luminance and color detail that SDR collapses.
| Method family | HDR input | Perceptual grounding | Reasoning | Continuous MOS | Interpretable | SROCC |
|---|---|---|---|---|---|---|
| Classical VQA (BRISQUE, VMAF) | ✗ | ✗ | ✗ | ✓ | ✗ | 0.41 |
| Deep VQA (FastVQA, DOVER) | ✗ | partial | ✗ | ✓ | ✗ | 0.51 |
| HDR-specific (HIDRO-VQA) Saini et al., WACV 2024 | ✓ | ✓ | ✗ | ✓ | ✗ | 0.85 |
| MLLM-VQA (Q-Insight, DeQA) | ✗ | ✗ | ✓ | partial | ✓ | 0.52 |
| HDR-Q (ours) | ✓ | ✓ | ✓ | ✓ | ✓ | 0.92 |
Only HDR-Q has HDR-aware perception, a continuous score and readable reasoning at once.
Blind to 10-bit PQ.
The MOS (mean opinion score) must come out of text.
GRPO (Group Relative Policy Optimization) ignores HDR tokens.
Adding HDR made it worse.

Scene categories across the 44K clips. The full figure is on a backup slide.
SigLIP-2 on native 10-bit PQ, trained to tell HDR from SDR.
HDR-Aware Policy Optimization makes the policy use the HDR tokens.
Ovis2.5 (9B), rank-4 LoRA. Reasoning in <think>, the MOS in <answer>.
SigLIP-2 finetuned on matched HDR-SDR pairs, with quality captions written by Qwen2.5-VL-72B.
HDR moves toward the caption, SDR moves away.
Input stays native 10-bit PQ.
HAPO rewards a different output for HDR than for SDR.
Ignore HDR and this KL is 0, so we maximize it.
Mutual information with HDR is bounded below.
| Term | What it does | Setting |
|---|---|---|
| Ã | Entropy-weighted advantage (HEW), new | λHEW = 0.3 |
| β DKL | Reference KL from GRPO, keeps training stable | β = 0.02 |
| + γ KHDR | Contrastive KL, new. Grounds the output in HDR | γ = 0.5 |
| − Hdual | Dual entropy, new. Prevents collapse | η1 = 0.01, η2 = 0.05 |
Training is two RL stages, modality alignment then score calibration, with no supervised fine-tuning (SFT).
| Method | SROCC ↑ | PLCC ↑ | RMSE ↓ |
|---|---|---|---|
| BRISQUE | 0.410 | 0.469 | 11.70 |
| CONTRIQUE | 0.625 | 0.605 | 15.02 |
| CONVIQT | 0.799 | 0.810 | 8.48 |
| HIDRO-VQA Saini et al., WACV 2024 | 0.851 | 0.878 | 6.09 |
| Q-Insight (best MLLM) | 0.517 | 0.562 | 20.78 |
| HDR-Q (SDR only) | 0.891 | 0.890 | 7.42 |
| HDR-Q (full) | 0.921 | 0.912 | 5.16 |
Zero-shot on LIVE-HDR and SFV+HDR, no fine-tuning on the target datasets.

Same video for both. HAPO (ours) is HDR-Q. Prompt and frame are on a backup slide.
Both get the same prompt, rate the video 0 to 100 and reason in <think> first.
Reports underexposure, color blooming and unnatural hues that are not in the video.
Highlights roll off without clipping, hues stay stable, no jitter or smear.
| Variant | SROCC | RMSE | Token entropy |
|---|---|---|---|
| GRPO baseline | 0.81 | 10.73 | 0.20 |
| + HDR encoder | 0.83 | 8.96 | 0.24 |
| HAPO without contrastive KL | 0.86 | 7.10 | 0.29 |
| HAPO without HEW | 0.88 | 6.11 | 0.27 |
| HAPO without dual entropy | 0.91 | 5.82 | 0.26 |
| HDR-Q (full) | 0.92 | 5.15 | 0.33 |
Token entropy rises from 0.20 to 0.33.
Thank you.
Backup
Full HAPO with γ = 0.5 on a curated Beyond8Bits subset of matched HDR-SDR pairs. The goal is to teach the model to attend to HDR-specific information.
The full Beyond8Bits corpus. The score reward is the main signal and γ is reduced. The goal is prediction accuracy while keeping the HDR grounding from stage 1.
Both stages use RL. There is no SFT stage.
Modality alignment and quality calibration pull against each other. Aggressive contrastive training hurts absolute accuracy, and pure quality training lets modality neglect back in. Doing them in sequence resolves the tension.
Setup: Ovis2.5 (9B) with rank-4 LoRA, K = 8, ε = 0.1, 4 H200 GPUs, BF16, T = 8 frames, native 10-bit PQ input.
Token entropy during training. GRPO collapses, HAPO keeps exploring.
Entropy drops, the policy turns deterministic, ignores the visual modality and shortcuts through text context.
Dual entropy holds H ≈ 0.33 at quality-critical positions, so exploration survives.
Chain of thought drops from 168 to 137 tokens (18% fewer). Boilerplate goes, the quality observations stay.
Dataset diversity, HDR against SDR characteristics, and HDR-Q performance gains.
Both get the same prompt, rate the video 0 to 100 and reason in <think> first.
Reports underexposure, color blooming and unnatural hues that are not in the video.
Highlights roll off without clipping, hues stay stable, no jitter or smear.