RAD: Refined Aesthetic Description Dataset

• The RAD dataset is a multi-dimensional resource featuring hierarchical descriptions that cover perception, cognition, and emotion in artistic images.
• It employs an automated, iterative LLM-driven pipeline to generate structured aesthetic comments and mitigate challenges like data scarcity.
• Empirical evaluations demonstrate improved performance in metrics such as SROCC and PLCC, supporting more accurate aesthetic assessments.

The Refined Aesthetic Description (RAD) dataset is a large-scale, multi-dimensional resource specifically designed for the quantitative assessment of artistic image aesthetics. It comprises 70 000 artwork images from established benchmarks—APDD, BAID, and VAPS—with each image paired to a detailed, hierarchical, machine-generated comment that describes its visual, semantic, and emotional attributes. RAD was constructed via an automated, iterative pipeline leveraging state-of-the-art LLMs, enabling generation of human-aligned, multidimensional aesthetic descriptions at scale. This dataset provides a solution to longstanding challenges of data scarcity, imbalance, and semantic fragmentation in AIGC aesthetics evaluation and is central to the ArtQuant framework's auxiliary description learning paradigm (Liu et al., 29 Dec 2025).

1. Dataset Architecture and Hierarchical Schema

Each RAD entry is represented by a quadruple $(x_v, s, D, y)$, where $x_v$ is a visual input (image identifier), $s \in [0,M]$ is a floating-point aesthetic score (Mean Opinion Score), $D \sim \mathcal{N}(\mu,\sigma^2)$ encodes the dataset-level score distribution, and $y$ is a block of hierarchical text. The annotation schema is structured into three explicit cognitive levels:

• Perception: Describes low-level visual features, such as color harmony, line quality, and composition balance.
• Cognition: Covers interpretive and art-historical elements, including subject matter, symbolism, and stylistic lineage.
• Emotion: Addresses affective response, describing mood, atmosphere, and emotional resonance elicited by the artwork.

No external ontology is formally integrated, but templates encode a controlled vocabulary for these attributes. This multi-level organization mirrors human aesthetic processing, ensuring coverage of surface and deep dimensions.

Field	Description	Example
Image ID	Visual Input ($x_v$)	"APDD_1234.jpg"
Score ($s$)	Aesthetic rating	4.2
Distribution	Dataset stats	$\mu=3.8$, $\sigma=0.7$
Description ($y$)	Hierarchical text	“Perception: … Cognition: … Emotion: …”

2. Automated Generation Pipeline

RAD is generated with a three-stage, LLM-driven loop, bypassing manual annotation:

1. Aesthetic Data Preprocessing: Original scores $s$ are normalized using dataset-level parameters $(\mu, \text{median}, \sigma)$ to correct for statistical bias. Prompts are structured to enforce the three-level description hierarchy.
2. Structured Description Generation: The generator $\mathcal{G}_{gen}$ (GPT-4o) receives $(x_v, P(s, \mu, \sigma, T))$ as input, where $T$ comprises aesthetic templates enforcing the Perception $\rightarrow$ Cognition $\rightarrow$ Emotion sequence. Output $y$ is a block with three marked sections.
3. Discriminative Quality Control: The discriminator $\mathcal{G}_{dis}$ (DeepSeek-chat) assesses alignment between generated $y$ and target score $s$ using an LLM-derived preference metric. Pairs below a consistency threshold are returned for re-generation until all entries meet alignment criteria.

This iterative process guarantees semantic coverage and efficient scaling, producing 70 000 descriptions at LLM speeds rather than costly human annotation.

3. Mathematical, Statistical, and Information-Theoretic Foundations

The RAD construction and ArtQuant supervision are grounded in explicit information-theoretic analysis:

• Notation:
  • $X$: Artwork images
  • $D$: Generated descriptions
  • $Z$: Latent multimodal representations
  • $Y$: Discrete score levels $\{l_1,\ldots,l_K\}$
• Entropy-minimization bounds:
  • Theorem 1: $H(Y|Z) \leq H(D|Z) + H(Y|D,Z)$
  • Conditional independence: If $Y \perp Z | D$, then $H(Y|Z) \leq H(D|Z) + H(Y|D)$
  • Error propagation ($\epsilon$-approximation):

  $$H(Y|Z) \leq H(Y|D) + \epsilon\log|Y| + H_2(\epsilon) + H(D|Z)$$ - $H(Y|D)$: Description sufficiency - $H(D|Z)$: Description generation ability

• Score-based distribution estimation:

  • Prediction: $x_{pred} = \sum_{i=1}^K p_i l_i$
  • Ground-truth distribution: $f = \text{Normal}(\mu, \sigma)$,

  $$f(x) = \frac{1}{\sigma\sqrt{2\pi}} \exp\left(-\frac{(x-\mu)^2}{2\sigma^2}\right)$$ - Estimation objective:

  $$\mu^*, \sigma^* = \arg\min \left\|\sum f(l_i) l_i - x_{gt}\right\|_2$$

  subject to $\sum f(l_i) = 1$

• Loss functions:

  • Description: $L_{CE} = -\sum_{t=1}^T\log P(t_t|t_{<t})$
  • Score regression: $L_{KL} = \sum p_i \log(p_i^{pred}/p_i)$
  • Combined: $L_{ASL} = L_{CE} + k L_{KL}$
  • Multi-task: $L_{MAT} = L_{CE} + k_{ASL} L_{ASL}$

This formulation formalizes the semantic adequacy and information gain from joint description learning.

4. Scalability, Data Quality, and Dimension Balance

The RAD pipeline scales to millions of images with negligible marginal cost, due to end-to-end automation by GPT-4o and DeepSeek-chat. Quality control is assured via strict score-comment alignment, using LLM-derived preference scores.

Empirical evaluations indicate improved annotation quality over manual approaches. For example, ArtQuant fine-tuned on RAD yields $+4.06\%$ SROCC (0.871 vs. 0.837) and $+3.47\%$ PLCC (0.894 vs. 0.864) on APDD compared to human comments. Ablation shows incremental benefits of full hierarchical descriptions:

Description Levels	SROCC	Improvement
Perception-only	0.865	+3.35%
+ Cognition	0.866	+3.46%
++ Emotion	0.871	+4.06%

This confirms balanced, multi-dimensional semantic content.

5. Integration with Aesthetics Assessment Models

RAD directly addresses data scarcity and dimension imbalance prevalent in prior AIGC aesthetics datasets, which overemphasized low-level perceptual features and suffered from sparse labels. The hierarchical schema prevents overfitting to superficial cues and supports a richer aesthetic representation.

When used to supervise ArtQuant, RAD enables state-of-the-art performance: on APDD, ArtQuant achieves SROCC/PLCC of 0.871/0.894 compared to ArtCLIP’s 0.810/0.840, requiring only $\sim$33% of the conventional training epochs. On BAID and VAPS benchmarks, ArtQuant converges in 2–8 epochs versus 15–200 for specialist baselines. Controlled ablation demonstrates that multi-task aesthetic training increases SROCC on small, specialized corpora by up to +14.7%, evidencing accelerated convergence and enhanced long-text understanding.

6. Context, Significance, and Future Directions

RAD introduces a scalable framework for collecting richly structured aesthetic commentary, leveraging LLMs to circumvent prohibitive annotation costs. Its three-level organization broadens cognitive and emotional coverage in image aesthetics, aligning machine assessment more closely with human perceptual and affective processes.

A plausible implication is that RAD’s paradigm can be adapted to even larger datasets or new domains (e.g., photographic, design, multimodal) as generalized LLMs continue to advance. The information-theoretic foundation, empirical metric gains, and efficient generation pipeline collectively position RAD as a reference resource for future research in artistic image evaluation, aesthetic regression, and multimodal understanding (Liu et al., 29 Dec 2025).