Multi-dimensional Quality Assessment

• Multi-dimensional quality assessment is a framework that decomposes quality into separate axes (e.g., brightness, noise, logical coherence) rather than relying on a single aggregate score.
• It leverages domain-specific metrics and structured observations to diagnose trade-offs and reveal hidden error sources across various applications.
• Fusion strategies and calibration techniques tailor the aggregation of independent quality signals for improved performance in areas like video enhancement, 3D rendering, and AI reasoning.

Multi-dimensional quality assessment is a family of evaluation frameworks that model quality as a structured set of dimensions rather than a single undifferentiated scalar. Across linked-data quality metadata, speech assessment, image and video quality assessment, text-to-3D evaluation, virtual try-on, surveillance facial analysis, dynamic 4D digital humans, decentralized LLM inference, and reasoning evaluation, the common premise is that a single score obscures distinct error sources, hides trade-offs, and can induce incorrect rankings under deployment-specific priorities (Debattista et al., 2014, Dong et al., 2023, Zhang et al., 2024, Şenol et al., 23 May 2026).

1. Conceptual foundations

In the surveyed literature, multi-dimensional quality assessment is consistently defined by decomposition. Instead of asking whether an artifact, sample, or system is simply “good” or “bad,” it asks along which axes quality is manifested, degraded, or preserved. In linked-data publishing, daQ represents dataset quality as multi-dimensional and statistical observations over metrics, computed resources, and timestamps, framed by the general principle that data quality is “fitness for use” (Debattista et al., 2014). In perceptual media, the same logic appears as explicit separation of spatial, temporal, semantic, geometric, aesthetic, or fidelity-related factors. In reasoning evaluation, the framework becomes behavioral: correctness, consistency, robustness, logical coherence, efficiency, and stability are treated as distinct signals rather than substitutes for one another (Şenol et al., 23 May 2026).

A central motivation is the insufficiency of single-number quality. Low-light video enhancement can simultaneously improve brightness while amplifying noise or inducing temporal brightness inconsistency, so a generic VQA score may miss the specific trade-off between restoration and artifact introduction (Dong et al., 2023). Text-to-3D assets may be well aligned to prompts yet geometrically defective, or geometrically plausible yet semantically incomplete (Fu et al., 24 Feb 2025). Surveillance facial images may look visually enhanced while losing identity-preserving fidelity, which is unacceptable for verification-centric use cases (Jiang et al., 7 Feb 2026). In LLM assessment, final-answer correctness can coexist with incoherent reasoning traces, and logical coherence is reported as orthogonal to correctness with $r=-0.172$ (Şenol et al., 23 May 2026).

This suggests that multi-dimensional quality assessment is not merely “more metrics,” but a change in evaluation ontology: quality becomes a vector, a profile, or a set of observations whose internal structure matters.

2. Domain-specific dimensional schemas

The dimensions chosen by each framework are domain-specific, but the pattern is stable: each dimension isolates a failure mode or perceptual factor that would be confounded in a single overall score.

Domain	Dimensions	Representative paper
Low-light video enhancement	brightness, noise, semantic spatial information; motion features and brightness consistency	Light-VQA (Dong et al., 2023)
UGC live video	semantic, distortion, motion	MD-VQA (Zhang et al., 2023)
Text-to-3D generation	alignment, geometry, texture, overall	HyperScore / MATE-3D (Zhang et al., 2024)
Virtual try-on	clothing fit, body compatibility, overall quality	VTONQA (Wei et al., 6 Jan 2026)
Surveillance facial images	noise, sharpness, colorfulness, contrast, fidelity, overall quality	SFIQA-Bench (Jiang et al., 7 Feb 2026)
Speech crowdsourcing	noisiness, coloration, discontinuity, loudness, reverberation, speech signal quality, overall quality	P.804-based framework (Naderi et al., 2023)
Reasoning quality in LLMs	correctness, consistency, robustness, logical coherence, efficiency, stability	Behavioral framework (Şenol et al., 23 May 2026)

Several taxonomies are especially characteristic. In low-light enhancement, Light-VQA isolates brightness and noise as the most impactful spatial factors and complements them with deep-learning-based semantic features; temporally, it combines motion features with handcrafted brightness consistency across frames (Dong et al., 2023). In text-to-3D, HyperScore treats alignment, geometry, texture, and overall quality as separate evaluation dimensions and conditions both feature fusion and score mapping on the target dimension (Zhang et al., 2024). In surveillance facial analysis, fidelity is elevated to a first-class dimension because enhancement may improve perceptual quality while damaging identity preservation (Jiang et al., 7 Feb 2026). In reasoning evaluation, efficiency and stability are explicitly separated from correctness, so a model can be accurate yet unstable or expensive (Şenol et al., 23 May 2026).

A different but structurally related schema appears in daQ, where quality is organized as Category–Dimension–Metric, and observations are represented with daq:metric, daq:computedOn, and daq:value inside a qb:DataSet (Debattista et al., 2014). Here, the “dimensions” are not perceptual attributes but analysis axes over which quality observations can be sliced and compared.

3. Datasets, subjective protocols, and annotation structure

A defining feature of multi-dimensional quality assessment is that annotation is dimension-wise rather than monolithic. The labels are usually collected independently for each dimension, then optionally fused into an overall score or used to train a unified predictor.

The low-light video benchmark behind Light-VQA contains 254 original low-light videos, enhanced by 8 low-light video enhancement algorithms to produce 2,060 videos in total (Dong et al., 2023). MATE-3D contains 1,280 textured meshes rendered from 160 prompts and 8 text-to-3D methods, with 107,520 ratings from 21 subjects across four dimensions using an 11-level impairment scale per ITU-T P.910 (Zhang et al., 2024). VTONQA contains 8,132 images from 11 VTON models with 24,396 mean opinion scores over clothing fit, body compatibility, and overall quality (Wei et al., 6 Jan 2026). SFIQA-Bench contains 5,004 surveillance facial images with six-dimensional ratings on a 5-point ACR scale from 100 participants (Jiang et al., 7 Feb 2026). UltraVQA scores videos on five dimensions, with each clip rated by at least 3 independent annotators and accompanied by GPT-generated rationales grounded in collective human judgments (Lin et al., 18 Feb 2026). MVQA-68K adds chain-of-thought-style explanations to seven-dimensional video quality labels (Pu et al., 15 Sep 2025).

Protocols are commonly derived from ITU recommendations. VTONQA applies ITU-style outlier handling and within-subject z-score normalization, computing

$$\mathrm{MOS}_j = \frac{1}{N_j}\sum_{i=1}^{N_j}\frac{(r_{ij}-\mu_i)/\sigma_i+3}{6}\times 100$$

after filtering (Wei et al., 6 Jan 2026). SFIQA-Bench uses BT.500-13 outlier processing but intentionally omits z-score normalization in the final stage to preserve the original 1–5 ACR semantics (Jiang et al., 7 Feb 2026). The multi-dimensional speech framework adapts P.804 and P.835 to crowdsourcing with staged qualification, gold questions, trapping questions, and looping playback (Naderi et al., 2023).

The literature also shows that data design itself becomes multi-dimensional. MATE-3D stratifies prompts into eight categories spanning single-object and multi-object scenarios (Zhang et al., 2024). UltraVQA includes fine-grained sub-attribute labels within each of five video dimensions (Lin et al., 18 Feb 2026). MVQA-68K combines seven dimension labels with causal explanations and a multi-prompt label vocabulary (Pu et al., 15 Sep 2025). This suggests that multi-dimensional assessment often requires richer annotation not only at the score level but also at the explanatory and categorical levels.

4. Modeling architectures and fusion strategies

Most learned systems follow a shared pattern: dimension-specific feature extraction, followed by fusion into either dimension-wise outputs, an overall score, or both. The main differences lie in how specialization and sharing are balanced.

Light-VQA is specialized to low-light video enhancement. Spatially, it handcrafts brightness and noise features and integrates them with deep-learning-based semantic features; temporally, it combines deep-learning-based motion features with handcrafted brightness consistency across frames, then fuses spatial and temporal information into a quality-aware representation (Dong et al., 2023). MD-VQA for UGC live video uses a no-reference design with semantic features from EfficientNetV2-S, handcrafted distortion descriptors, and motion features from ResNet3D-18, plus explicit temporal fluctuation modeling through adjacent-frame absolute differences (Zhang et al., 2023).

HyperScore for text-to-3D introduces a more explicitly conditional formulation. It uses CLIP visual and textual encoders, learns dimension condition features from the meta texts “alignment quality,” “geometry quality,” “texture quality,” and “overall quality,” and computes

$\hat{q}_i=\psi\Big(\phi(x,t)\mid \pi(f_c^i)\Big),$

where a hypernetwork $\pi$ generates dimension-specific mapping parameters for the shared head $\psi$ (Zhang et al., 2024). This makes the evaluator dimension-aware not only at the feature stage but also at the scoring stage.

Multi-task designs are common in image quality assessment. MDIQA uses two shared backbones—one for technical dimensions and one for aesthetic dimensions—then assigns lightweight independent heads to each dimension and learns an image-adaptive weight vector before overall fusion by a 3-layer MLP (Yao et al., 23 Aug 2025). LEIQ-Assessor uses a pre-trained SigLIP2 Vision Transformer with seven task-specific MLP regressors, jointly predicting overall MOS and six perceptual sub-attributes for low-light enhanced images (Sun et al., 29 Jun 2026). SFIQA-Assessor uses three facial views, low-rank cross-view feature interaction, and six learnable task tokens for unified multi-task regression (Jiang et al., 7 Feb 2026).

Large multimodal models extend the same logic to token-level fusion. AIGVEval decomposes AI-generated video quality into technical quality, motion quality, and video semantics, encodes each dimension separately, and injects the resulting tokens into a Vicuna v1.5–7B LLM using semantic anchors such as “the technical quality features of the video are” (Qi et al., 5 Jun 2025). DynaMesh-Rater for dynamic 4D human meshes concatenates projected visual tokens, motion tokens from SlowFast, and geometry tokens derived from dihedral-angle statistics, then uses an LMM with LoRA-based instruction tuning to regress quality (Li et al., 4 Oct 2025).

Not all frameworks are neural. daQ maps quality observations to the W3C Data Cube vocabulary, enabling multidimensional slicing, querying, and visualization without a learned regressor (Debattista et al., 2014). In this form, multi-dimensional assessment is a representational framework rather than a prediction model.

5. Scoring, calibration, and interpretability

The output of a multi-dimensional assessor can be a set of independent scores, an overall score, or both. The way these outputs are calibrated and aggregated is itself a major design variable.

A common starting point is mean opinion score aggregation: $\mathrm{MOS}_i=\frac{1}{N}\sum_{j=1}^{N}s_{ij},$ used explicitly in surveillance facial assessment and many other subjective protocols (Jiang et al., 7 Feb 2026). Objective prediction is then usually evaluated with PLCC and SRCC, and often KRCC and RMSE. HyperScore reports PLCC, SRCC, and KRCC with logistic mapping before PLCC computation (Zhang et al., 2024). T23DAQA applies a 5-parameter logistic mapping before PLCC on its three dimensions of quality, authenticity, and text-asset correspondence (Fu et al., 24 Feb 2025). LEIQ-Assessor instead computes PLCC and SRCC directly on raw predictions without logistic calibration (Sun et al., 29 Jun 2026).

Aggregation can be fixed, learned, or deployment-specific. SA-IQA learns a linear fusion $Score=x^\top w^*$ from pairwise human preference data using a Bradley–Terry loss, and reports that optimal weighting outperforms equal weighting in spatial aesthetics assessment (Gao et al., 4 Dec 2025). The reasoning-quality framework defines

$Q_{bal}(m)=\frac{1}{6}\sum_d S_m^{(d)}$

for balanced scoring and

$Q_w(m)=\sum_d w_d S_m^{(d)}$

for deployment-weighted scoring, showing that model rankings invert under different weight vectors (Şenol et al., 23 May 2026). In decentralized LLM inference, a composite quality signal

$S(q,y)=\sum_{d\in D} w_d\,\tilde{s}^{(d)}$

is explicitly pruned and renormalized when some dimensions are unreliable or negatively correlated with reference quality (Tian et al., 4 Mar 2026).

Interpretability is often the principal justification for multi-dimensionality. P.804-style speech assessment exposes whether degradation is driven by noisiness, coloration, discontinuity, loudness, or reverberation rather than by a single overall MOS (Naderi et al., 2023). UltraVQA supplements five-dimensional video ratings with GPT-generated rationales grounded in human tags (Lin et al., 18 Feb 2026). MVQA-68K adds causally oriented chain-of-thought explanations to seven video dimensions (Pu et al., 15 Sep 2025). In LLM reasoning evaluation, dimension-level profiles directly reveal whether a model fails on logical coherence despite correct answers (Şenol et al., 23 May 2026).

A recurrent caution in the literature is that dimensions are not automatically valid merely because they are interpretable. In decentralized LLM inference, seemingly reasonable dimensions such as query-output alignment or agreement/uncertainty can be task-dependent and even negatively correlated with reference quality unless they are calibrated and, when necessary, pruned (Tian et al., 4 Mar 2026).

6. Applications, misconceptions, and future directions

The applications of multi-dimensional quality assessment are broad but structurally similar: model benchmarking, failure diagnosis, system optimization, and policy or reward design. Light-VQA is designed to evaluate low-light video enhancement outputs rather than generic video quality (Dong et al., 2023). MDIQA can be deployed as a flexible perceptual loss for image restoration, with dimension weights adjusted to user preference (Yao et al., 23 Aug 2025). SA-IQA serves both as a Best-of-$$\mathrm{MOS}_j = \frac{1}{N_j}\sum_{i=1}^{N_j}\frac{(r_{ij}-\mu_i)/\sigma_i+3}{6}\times 100$$0 selector and as a reward signal in GRPO-based optimization of image generation pipelines (Gao et al., 4 Dec 2025). T23DAQA supports text-to-3D asset selection and prompt engineering feedback (Fu et al., 24 Feb 2025). In decentralized LLM inference, multi-dimensional scoring is integrated into Proof of Quality as a cost-aware reward signal (Tian et al., 4 Mar 2026). In linked-data publishing, daQ supports browsing datasets by quality, version analysis, and link identification (Debattista et al., 2014).

A common misconception is that a higher-dimensional evaluator is necessarily superior. The evidence is more conditional. UltraVQA and MVQA-68K show that multi-dimensional labels can improve performance and interpretability in video quality modeling (Lin et al., 18 Feb 2026, Pu et al., 15 Sep 2025). However, the decentralized inference study shows that an uncalibrated default composite can underperform a strong single semantic evaluator, and only matches or exceeds it after unreliable dimensions are removed and weights are renormalized (Tian et al., 4 Mar 2026). This suggests that dimensional proliferation without auditing can degrade, rather than improve, evaluation quality.

Another misconception is that “overall quality” subsumes the constituent dimensions. Several studies directly challenge this. In surveillance facial images, overall quality is strongly explained by fidelity and sharpness, with a reported linear regression

$$\mathrm{MOS}_j = \frac{1}{N_j}\sum_{i=1}^{N_j}\frac{(r_{ij}-\mu_i)/\sigma_i+3}{6}\times 100$$1

$$\mathrm{MOS}_j = \frac{1}{N_j}\sum_{i=1}^{N_j}\frac{(r_{ij}-\mu_i)/\sigma_i+3}{6}\times 100$$2, which shows unequal dimensional contributions (Jiang et al., 7 Feb 2026). In VTONQA, body compatibility tends to be higher than clothing fit, and overall quality appears more heavily influenced by body compatibility than by clothing fit under current model behaviors (Wei et al., 6 Jan 2026). In reasoning evaluation, logical coherence cannot be inferred from correctness (Şenol et al., 23 May 2026).

Future directions described in the literature are largely expansions of scope rather than repudiations of the paradigm. HyperScore explicitly proposes extending text-to-3D evaluation to physical plausibility, lighting/material consistency, multi-view consistency, and personalized preference modeling (Zhang et al., 2024). LEIQ-Assessor identifies explicit modeling of inter-attribute dependencies and temporal extension to low-light video enhancement as promising directions (Sun et al., 29 Jun 2026). daQ emphasizes custom metric extension, provenance, and integration with DCAT, VoID, and PROV-O (Debattista et al., 2014). This suggests a likely trajectory in which multi-dimensional quality assessment becomes increasingly modular, domain-specific, and explanation-aware, while retaining the core principle that quality must be represented as a structured profile rather than a single undifferentiated number.