CompressedVQA-HDR: Unified HDR Quality Model

• The paper introduces a dual-branch framework that combines Swin Transformer-based full-reference and SigLip 2-based no-reference modules to assess compressed HDR video quality.
• It leverages cross-dataset generalization through SDR pre-training, mixed-dataset training, and HDR fine-tuning to address the scarcity of HDR labels.
• Experiments demonstrate significant performance improvements over traditional methods, with superior SRCC and PLCC metrics on HDRSDR-VQA and ICME challenge benchmarks.

Searching arXiv for CompressedVQA-HDR and closely related HDR VQA work to ground the article in current papers. CompressedVQA-HDR is a generalized framework for quality assessment of compressed high dynamic range video in both the full-reference and no-reference settings. It was introduced as a response to the limited transferability of conventional compressed-video quality models to HDR content, whose higher dynamic range, increased bit depth, and different signal characteristics alter both the visibility and the semantics of distortions. The framework couples a Swin Transformer-based full-reference branch with a SigLip 2-based no-reference branch, and emphasizes cross-dataset generalization through SDR pre-training, mixed-dataset training, and HDR fine-tuning. In the reported ICME 2025 Grand Challenge, CompressedVQA-HDR-FR ranked first in the full-reference track (Sun et al., 16 Jul 2025).

1. Position within HDR video quality assessment

CompressedVQA-HDR emerged within a research trajectory that moved from small controlled HDR databases toward broader, more realistic, and more heterogeneous benchmarks. Early large-scale subjective HDR video work established that HDR10 quality prediction cannot be reduced to SDR assumptions, and that compression and resolution changes are central HDR delivery impairments; the LIVE-HDR study, for example, released 310 HDR videos derived from 31 source contents and more than 20,000 opinion scores under dark and bright ambient conditions (Shang et al., 2022). Subsequent work showed that HDR-versus-SDR preference depends jointly on display technology, bitrate, and scaling, rather than on dynamic range format alone (Ebenezer et al., 2023). Later datasets expanded the distortion manifold toward user-generated and platform-like scenarios: CHUG introduced 856 UGC-HDR source videos and 5,992 total videos through a bitrate-ladder design, while Beyond8Bits scaled to approximately 44K HDR-UGC videos with more than 1.5 million crowd ratings and an explicitly HDR-aware no-reference model (Saini et al., 10 Oct 2025, Saini et al., 1 Mar 2026).

Within this landscape, CompressedVQA-HDR addresses a specific gap: a unified deep framework for compressed HDR assessment that targets both full-reference and no-reference operation while foregrounding generalization. This positioning differs from earlier HDR-specific approaches that were either feature augmentations for existing SDR models, as in HDRMAX (Ebenezer et al., 2023), or methods optimized for a single public HDR benchmark, as in HIDRO-VQA (Saini et al., 2023). A plausible implication is that CompressedVQA-HDR should be understood less as a narrowly specialized HDR psychovisual model than as a generalizable compressed-VQA adaptation layer for HDR content.

2. Architectural formulation

The framework starts from temporally subsampled video. Given distorted and reference videos

$$\mathbf{x}^d = \{\mathbf{x}^d_i\}_{i=0}^{N-1}, \qquad \mathbf{x}^r = \{\mathbf{x}^r_i\}_{i=0}^{N-1},$$

with frames in $\mathbb{R}^{H\times W\times 3}$ and frame rate $R$, the videos are reduced to 1 fps:

$$K = \left\lfloor \frac{N}{R} \right\rfloor, \qquad \mathbf{z}^d_i = \mathbf{x}^d_{\lfloor R \times i \rfloor}, \qquad \mathbf{z}^r_i = \mathbf{x}^r_{\lfloor R \times i \rfloor}.$$

Each sampled frame is then resized by bicubic interpolation to $384 \times 384$ (Sun et al., 16 Jul 2025).

In the full-reference branch, each paired reference and distorted frame is passed through a Swin Transformer backbone. If

$$\mathbf{f}^d_i = \mathrm{Swin}(\mathbf{z}^d_i), \qquad \mathbf{f}^r_i = \mathrm{Swin}(\mathbf{z}^r_i),$$

then intermediate-layer features from all Swin stages are used to compute two DISTS-style descriptors: deep textural similarity and deep structural similarity. These are defined as

$$\mathbf{f}_{T}\!\left(\mathbf{f}^r_i, \mathbf{f}^d_i\right) = \frac{2 \,\boldsymbol{\mu}_{\mathbf{f}^r_i}\,\boldsymbol{\mu}_{\mathbf{f}^d_i} + c_1} {(\boldsymbol{\mu}_{\mathbf{f}^r_i})^2 + (\boldsymbol{\mu}_{\mathbf{f}^d_i})^2 + c_1},$$

and

$$\mathbf{f}_{S}\!\left(\mathbf{f}^r_i, \mathbf{f}^d_i\right) = \frac{2 \,\boldsymbol{\sigma}_{\mathbf{f}^r_i \mathbf{f}^d_i} + c_2} {(\boldsymbol{\sigma}_{\mathbf{f}^r_i})^2 + (\boldsymbol{\sigma}_{\mathbf{f}^d_i})^2 + c_2},$$

where $\boldsymbol{\mu}$ denotes global means, $\boldsymbol{\sigma}^2$ global variances, $\mathbb{R}^{H\times W\times 3}$0 global covariance, and $\mathbb{R}^{H\times W\times 3}$1 are stabilizing constants. The structure and texture similarities from each Swin stage are concatenated into a frame-level full-reference feature representation.

In the no-reference branch, the distorted frame alone is processed by SigLip 2. The reported representation is the global mean of the final-layer feature maps from the SigLip 2 visual encoder, yielding a frame-level quality-aware descriptor. This “minimalistic” design omits hand-crafted HDR modules and instead relies on the final pretrained visual representation as the quality signal.

Both branches use the same regressor structure. A two-layer MLP with 128 hidden neurons and a single output predicts a frame-level quality score,

$\mathbb{R}^{H\times W\times 3}$2

and the video-level score is the arithmetic mean over sampled frames,

$\mathbb{R}^{H\times W\times 3}$3

The paper does not introduce an explicit temporal model beyond 1 fps sampling and score averaging, so motion is handled only indirectly through frame content and temporal sparsity (Sun et al., 16 Jul 2025).

3. Training strategy and generalization mechanism

The central training problem identified by CompressedVQA-HDR is the scarcity of HDR labels. The full-reference and no-reference branches address this differently. For the full-reference model, the authors use transfer learning: pre-train on the Compressed UGC VQA dataset from the ICME 2021 challenge, then fine-tune on HDRSDR-VQA. In the reported setup, Compressed UGC VQA contains 6,400 training clips and 800 validation clips, with each reference compressed into seven distorted versions using H.264/AVC at various CRF levels (Sun et al., 16 Jul 2025).

The no-reference branch uses iterative mixed-dataset training, abbreviated IMDT. A shared feature extractor with parameters $\mathbb{R}^{H\times W\times 3}$4 is trained jointly with dataset-specific regressors $\mathbb{R}^{H\times W\times 3}$5 so that heterogeneous datasets can contribute without forcing a common score scale. For dataset

$\mathbb{R}^{H\times W\times 3}$6

the loss is

$\mathbb{R}^{H\times W\times 3}$7

Optimization proceeds sequentially across datasets,

$\mathbb{R}^{H\times W\times 3}$8

To balance differently sized datasets, the number of epochs for dataset $\mathbb{R}^{H\times W\times 3}$9 in one loop is

$R$0

with $R$1 in the reported experiments. After mixed-dataset training, the no-reference model is fine-tuned on HDRSDR-VQA (Sun et al., 16 Jul 2025).

The IMDT pool comprises HDRSDR-VQA, Compressed UGC VQA, WaterlooIVC-4K, ITM-HDR-VQA, and LIVE Livestream. WaterlooIVC-4K contributes 1,200 compressed sequences from 20 pristine 10-second 4K sources, encoded by AVC, HEVC, VP9, AVS2, and AV1 across three resolutions and four distortion levels. ITM-HDR-VQA contributes 200 inverse tone-mapped HDR videos from 20 scenes. LIVE Livestream contributes 315 videos from 45 high-motion sports contents with compression, aliasing, flicker, judder, frame drops, and interlacing (Sun et al., 16 Jul 2025).

Optimization details are specified but not exhaustively. Both branches use Adam. The initial learning rate is $R$2 for the full-reference model and $R$3 for the no-reference model, batch size is 6, full-reference pre-training lasts 10 epochs, HDR fine-tuning lasts 30 epochs, IMDT uses 3 loops, and inputs are resized to $R$4. The paper states that a PLCC loss is used for both branches, but does not print its explicit analytic form (Sun et al., 16 Jul 2025).

4. Experimental basis and benchmark setting

The reported evaluation centers on HDRSDR-VQA. In the CompressedVQA-HDR experiments, HDRSDR-VQA is described as containing 360 videos, split evenly into 180 HDR and 180 SDR sequences derived from 20 open-source HDR source videos, with an 8:2 split by reference video into training and validation sets (Sun et al., 16 Jul 2025). The original HDRSDR-VQA dataset paper, by contrast, introduced a larger paired-format benchmark with 960 total videos from 54 source sequences, 480 HDR10 and 480 SDR, eight bitrate-resolution distortion levels plus reference, pairwise comparisons from 145 participants, and JOD scaling on six consumer HDR televisions (Chen et al., 27 May 2025). This suggests that CompressedVQA-HDR operates on a reduced experimental configuration of that benchmark, although the paper itself does not formalize the relation.

Performance is reported with SRCC, KRCC, PLCC, and RMSE. In the full-reference setting, the comparison set includes SSIM, MS-SSIM, LPIPS, DISTS, VMAF, and CompressedVQA. In the no-reference setting, the baselines include NIQE, SimpleVQA, FAST-VQA, DOVER, MinimalisticVQA, and CompressedVQA. The challenge evaluation further compares against PSNR-Y and VMAF on the held-out ICME 2025 test set (Sun et al., 16 Jul 2025).

This benchmark design places CompressedVQA-HDR at the intersection of paired HDR/SDR evaluation and compressed-video generalization. A plausible implication is that the framework is intended not only for native HDR-only quality prediction, but also for the mixed HDR/SDR decision surface emphasized by HDRSDR-VQA.

5. Reported performance

On the HDRSDR-VQA validation set, the full-reference model substantially exceeds all listed baselines. The reported results are SRCC 0.9197, PLCC 0.9348, KRCC 0.7498, and RMSE 0.4489. For comparison, CompressedVQA reaches SRCC 0.7814 and PLCC 0.7754; VMAF reaches SRCC 0.7358 and PLCC 0.7081; DISTS reaches SRCC 0.6729 and PLCC 0.6546 (Sun et al., 16 Jul 2025).

The no-reference model also leads the listed baselines on HDRSDR-VQA. Its reported performance is SRCC 0.9241, PLCC 0.9261, KRCC 0.7663, and RMSE 0.4768. The strongest competing baseline in SRCC is MinimalisticVQA at 0.8810; SimpleVQA attains PLCC 0.9039 but lower SRCC 0.8020; DOVER reports SRCC 0.8111 and PLCC 0.7985; NIQE fails badly, with SRCC 0.0519 and PLCC 0.0103 (Sun et al., 16 Jul 2025).

On the ICME 2025 challenge test set, the full-reference branch reports SRCC 0.932, PLCC 0.941, KRCC 0.778, and RMSE 0.482. The corresponding VMAF numbers are SRCC 0.905, PLCC 0.901, KRCC 0.725, and RMSE 0.597, while PSNR-Y reports SRCC 0.674 and PLCC 0.677. The paper states that the proposed full-reference model achieves a 2.98% SRCC gain over VMAF and ranked first in the challenge’s full-reference track (Sun et al., 16 Jul 2025).

The ablations isolate two sources of improvement. In the full-reference branch, replacing ResNet-50 with Swin-B improves performance from SRCC 0.8864 and PLCC 0.8579 to SRCC 0.9011 and PLCC 0.9029, and adding SDR pre-training further improves to SRCC 0.9197 and PLCC 0.9348. In the no-reference branch, the progression from ResNet-50 to Swin-B to SigLip-B v2 yields SRCC 0.8473, 0.8763, and 0.8924 respectively, while adding IMDT raises performance to SRCC 0.9241 and PLCC 0.9261. The ablation results therefore show that backbone strength and cross-dataset transfer are both essential components of the final system (Sun et al., 16 Jul 2025).

6. Relation to prior methods, interpretation, and limitations

CompressedVQA-HDR belongs to a broader sequence of attempts to adapt quality assessment to HDR. HDRMAX proposed a feature augmentation that makes SDR-designed models more sensitive to bit depth and HDR-visible distortions in locally bright and dark regions; on LIVE HDR, VMAF+HDRMAX reached SRCC 0.8528 in the full-reference setting, while VBLIINDS+HDRMAX+NOISE reached SRCC 0.8492 in the no-reference setting (Ebenezer et al., 2023). HIDRO-VQA instead transferred self-supervised SDR representations to HDR using unlabeled HDR videos and reached SROCC 0.8793 in dark ambient and 0.8930 in bright ambient on LIVE-HDR (Saini et al., 2023). In the full-reference compressed-HDR regime, 3C-FUNQUE+ and its HDRMAX-augmented variants achieved strong LIVE-HDR performance, with dark-ambient SROCC up to 0.9022 and bright-ambient SROCC up to 0.8906 (Venkataramanan et al., 2023).

Relative to these methods, CompressedVQA-HDR replaces explicit HDR handcrafting with large pretrained backbones and generalization-oriented training. This suggests that its main novelty lies less in a new HDR psychovisual mechanism than in the way representation learning, dataset transfer, and dual FR/NR formulation are combined. The paper’s own limitations reinforce that reading. It does not introduce explicit temporal modeling beyond frame averaging, does not report parameter counts or runtime, and does not add bespoke luminance-adaptation or transfer-function-aware modules. Its HDR specificity is therefore largely empirical and data-driven rather than mechanistically modeled (Sun et al., 16 Jul 2025).

A broader field-level implication is that CompressedVQA-HDR marks a transitional stage between earlier lab-oriented HDR benchmarks and later large-scale HDR-UGC systems. Later work such as Beyond8Bits and HDR-Q extends the same generalization agenda toward tens of thousands of HDR-UGC videos, HDR-native multimodal reasoning, and explicit HDR-vs-SDR contrastive learning (Saini et al., 1 Mar 2026). In that sense, CompressedVQA-HDR can be read as a foundation-model adaptation of compressed VQA to HDR, positioned between feature-engineered HDR extensions and the subsequent move toward large-scale, UGC-driven, HDR-native learning.