Wide-Spectrum Aesthetics Dataset

• Wide-spectrum aesthetics datasets are a design principle that emphasizes breadth through diverse rating scales, multiple human ratings, and heterogeneous visual domains.
• They preserve variation in aesthetic judgment by including both attractive and unattractive images, detailed annotation methods, and a range of perception and production tasks.
• Comparative analyses reveal that differences in image genres, rating constructs, and evaluation strategies challenge the transferability of results across datasets.

A wide-spectrum aesthetics dataset is best understood not as a single benchmark name but as a dataset design principle: breadth may come from wide score ranges, multiple human ratings per image, heterogeneous visual domains, fine-grained aesthetic dimensions, production-oriented tasks, or prompt-conditioned pairwise preferences rather than a single binary beauty label. Comparative work across twelve aesthetics datasets shows that the field is structurally fragmented across photographs, paintings, AI-generated images, beauty/liking/aesthetic-quality constructs, and different rating populations, and that results do not transfer uniformly across datasets (Bartho et al., 2023). This suggests that “wide-spectrum” in aesthetics refers less to raw scale than to preserving variation in aesthetic judgment, image genre, and evaluation form; later datasets make that breadth explicit by modeling disagreement, including ugly images, spanning art and design domains, or expanding assessment from perception to production (Wang et al., 2016, Parraga et al., 8 May 2025, Wang et al., 9 Nov 2025).

1. Breadth as a dataset design principle

The literature operationalizes breadth along several axes. One axis is rating range: some datasets attempt to cover the full continuum from very ugly to very beautiful rather than concentrating around attractive images. A second axis is domain range: some datasets span photographs, paintings, calligraphy, design, 3D renderings, or AI-generated images rather than one medium alone. A third axis is annotation structure: aesthetics may be encoded as scalar scores, rating distributions, pairwise preferences, question answering, grounded localization, or expert textual critique. A fourth axis is task range: newer benchmarks evaluate not only perception but also editing, image selection, framing, and other production-oriented judgments (Bartho et al., 2023, Wang et al., 9 Nov 2025).

A recurring misconception is that scale alone guarantees generality. The comparative survey of twelve datasets instead shows that image genre, rating construct, collection method, and rating population materially affect what is learnable and how reproducible results are across datasets (Bartho et al., 2023). Another misconception is that all “aesthetics datasets” use the same target. Some are direct human-rating corpora, some are pairwise preference corpora, and some are curated style or appearance resources whose aesthetic relevance comes from downstream use rather than explicit beauty judgments. This suggests that dataset choice should follow the target construct: visual beauty, preference under a prompt, expert critique, design diagnosis, or physiologically grounded appearance analysis.

Resource	Breadth emphasized	Supervision
AVA (Wang et al., 2016)	broad photographic content and inter-rater disagreement	1–10 ratings, style subset, binary/mean/distribution tasks
MSC (Parraga et al., 8 May 2025)	balanced ugly-to-beautiful range with minimal semantic content	100 ratings per image, truncated-Gaussian mean and SD
BAID (Yi et al., 2023)	artworks across various art forms	vote-derived score over 60,337 artistic images
HumanBeauty (Liao et al., 31 Mar 2025)	human-image aesthetics across 12 dimensions	MOS for overall and fine-grained HIAA
AesEval-Bench (An et al., 1 Mar 2026)	graphic-design principles across tasks	aesthetic judgment, region selection, precise localization
HPDv3 (Ma et al., 5 Aug 2025)	low-to-high quality real and generated prompt-conditioned images	1.17M pairwise human preferences

2. Photographic foundations: AVA and distributional supervision

The modern wide-spectrum discussion is anchored by AVA, which is described as “a large-scale collection of images and meta-data derived from DPChallenge.com,” containing “over 250,000 images with aesthetic ratings from 1 to 10,” plus “a 14,079 subset with binary style labels.” Those style labels cover fourteen photographic attributes, including Complementary Colors, High Dynamic Range, Macro, Shallow DOF, Rule of Thirds, Silhouettes, and Vanishing Point. Each image has multiple human ratings on the discrete 1–10 scale, which enables three task formulations: binary classification by thresholding the mean score, scalar score prediction, and rating-distribution prediction. The dataset is especially important because it preserves disagreement: the paper stresses that two images with equal mean score may have very different deviations among raters, and cites prior AVA analysis that rating distributions are “largely Gaussian” for 99.77% of images (Wang et al., 2016).

AVA’s breadth is not only semantic but also photographic and technical. Another AVA study describes it as containing about 250,000 images from dpchallenge.com, each rated by an average of about 210 users, with more than 25,000 unique resolutions, average image size about $629 \times 497$, and maximum size up to $800 \times 800$. That work argues that aesthetics prediction should be formulated over the full score range, not only a binary high/low split, because cropping, down-scaling, or warping can suppress composition, blur, sharpness, and other qualities that human raters actually judged in the original images (Hosu et al., 2019).

At the same time, AVA illustrates a structural caveat of wide-spectrum photographic data. Because it comes from DPChallenge, ratings are not purely intrinsic judgments detached from social context; challenge topics and comments can influence the scores, and the paper gives examples where creativity or mismatch to a challenge theme affected ratings (Wang et al., 2016). A plausible implication is that AVA is broad precisely because it mixes composition, technical quality, semantics, novelty, and community context, but that same mixture complicates any attempt to interpret it as a pure measure of isolated visual beauty.

3. Extending the range: ugliness, disagreement, and ranking

A major criticism of traditional aesthetics corpora is beauty bias. The Minimum Semantic Content (MSC) database was created to address that bias directly by restricting images to natural scenes and natural objects while excluding people, animals, and human-made objects, and by generating ugly images through explicit visual manipulations. The final dataset contains 10,426 images: 5,684 original unmodified images, 919 beautified images, 872 uglified images, and 2,951 auto-uglified images. Each image received 100 human ratings on a 5-point Likert scale from “very ugly” to “very beautiful,” and the authors fit a truncated Gaussian to each rating histogram, extracting a fitted mean $\mu$ and standard deviation as the final continuous representation. The paper explicitly states that it “managed to remove the bias towards high ratings, creating a more uniform distribution,” and shows that including ugly images can modify or even invert observed relations between image features and aesthetic valuation (Parraga et al., 8 May 2025).

A second way to widen the spectrum is to replace absolute labels with relative ordering. “A Computational Approach to Relative Aesthetics” constructs a 43,000-pair dataset from AVA, split into 20,000 training pairs, 3,000 validation pairs, and 20,000 test pairs. Pairs are retained only if the absolute difference between average ratings is at least 1, both images have rating variance below 2.6, and both belong to the same category. The stated rationale is that no single threshold can correctly determine ranking order across the whole dataset: some pairs contain two images that would both be “beautiful” under a binary split, while others contain two images that would both be “non-beautiful” (Chandakkar et al., 2017).

These two datasets widen aesthetics in different directions. MSC widens the score continuum by deliberately populating the ugly end; the AVA-derived pair corpus widens the decision structure by treating aesthetics as comparative ranking rather than thresholded classification. This suggests that “wide-spectrum” can mean either broader coverage of aesthetic values or a richer representation of ordinal structure, and that both are distinct from merely increasing image count.

4. Domain-specific expansions in art, human imagery, and expert critique

Several recent datasets are broad not across all visual media, but within a specialized domain. For artistic-image aesthetics, BAID introduces 60,337 artistic images with more than 360,000 votes from online users. It is described as “constructed completely from artworks,” covering “various art forms,” “a wide range of artistic styles and painting themes,” and source contests that do not limit subject matter, style, or medium. Its supervision remains scalar and vote-derived rather than distributional, but it substantially broadens artistic coverage relative to earlier small art-aesthetics datasets (Yi et al., 2023).

For human-centered imagery, HumanBeauty is presented as the first dataset purpose-built for Human Image Aesthetic Assessment. It contains 108,586 human images in total, comprising HumanBeauty-58K with 58,564 images filtered from public datasets and HumanBeauty-50K with 50,022 Internet-collected images. The full dataset is split into 94,099 training images and 14,487 test images. Its key contribution is a hierarchical 12-dimensional standard spanning facial aesthetics, general appearance aesthetics, environment, and overall aesthetic, with each image in the 50K subset annotated by at least 9 raters and scored in the continuous range $[0,1]$ via Mean Opinion Score (Liao et al., 31 Mar 2025).

Other datasets broaden aesthetics by adding structured explanation rather than only new image domains. CompArt extends a WikiArt-based corpus to 80,032 artworks spanning 1,119 artists and 27 art styles, and adds a concise caption, top-3 predicted styles, and Principles of Art analyses over ten dimensions: Balance, Harmony, Variety, Unity, Contrast, Emphasis, Proportion, Movement, Rhythm, and Pattern. The annotations are produced by GPT-4o in a constrained JSON format with prominence levels and analyses, and the release is explicitly intended to support user-specified compositional control rather than a universal scalar beauty target (Jin et al., 15 Mar 2025). RAD similarly reframes artistic aesthetics as a 70k, multi-dimensional structured description dataset organized around perception, cognition, and emotion. It is generated with GPT-4o and checked by DeepSeek-chat, and the paper explicitly states that historical background and cultural significance are presently omitted (Liu et al., 29 Dec 2025).

ArtiMuse-10K occupies an intermediate position between specialization and breadth. It contains 10,000 images spanning 5 main categories and 15 subcategories—Graphic Design, 3D Design, AIGC, Photography, and Painting & Calligraphy—with each image annotated by professional experts using 8-dimensional attribute analysis and an overall score. The appendix gives a 9,000/1,000 train/test split and subcategory counts such as Daily Photo, Photographic Art, Architecture, Portrait, Movie still, Digital Art, Chinese Painting, Product Design, Sculpture, and Graphic Design. This makes the dataset unusually broad across media while retaining expert critique as its central supervision format (Cao et al., 19 Jul 2025).

5. Multimodal, design, video, and prompt-conditioned benchmarks

A major recent trend is the movement from scalar image labels to multimodal and task-structured supervision. AesVQA is the first aesthetic visual question answering dataset for photographs. It contains 72,168 high-quality images from Flickr and 324,756 aesthetic question-answer pairs, with train/validation/test splits of 58,168, 7,000, and 7,000 images. Its labels cover composition, color, subject, lighting, genres, techniques, and emotions, and include both objective QA pairs generated by aesthetic-attribute analysis algorithms and subjective QA pairs derived from numerical labels and sentiment analysis (Jin et al., 2022).

For graphic design, AesEval-Bench introduces a benchmark spanning four dimensions—layout, font/typography, color, and graphics—twelve indicators, and three fully quantifiable tasks: aesthetic judgment, region selection, and precise localization. Its companion training set, AesEval-Train, contains 30k question-answer pairs with indicator-grounded reasoning paths. The benchmark is described as graphic-design-specific rather than general-purpose, but it is unusually broad within that domain because it combines formal design principles with spatial grounding and localization (An et al., 1 Mar 2026).

For video, AesVideo-Bench extends breadth into time and cinematic structure. It uses three top-level dimensions—Visual Aesthetics, Visual Fidelity, and Visual Plausibility—and decomposes them into fifteen criteria such as Color Quality, Light Direction, Structural Stability, Shot Composition, Focal Length, and Camera Angle. The held-out benchmark contains about 2.5K video pairs, with the appendix giving 2,639 samples, and it sits inside a larger preference-collection pipeline of about 20K high-quality samples used for reward-model training (Han et al., 30 Apr 2026). AesTest broadens aesthetics along another axis: capability coverage. It contains 17,885 images and 8,757 QA items organized into ten sub-tasks across Perception, Appreciation, Creation, and Photography, thereby evaluating not only understanding but also critique, image selection, retouch preference, and framing choice (Wang et al., 9 Nov 2025).

The most expansive recent notion of wide-spectrum supervision appears in HPDv3, the dataset underlying HPSv3. HPDv3 contains 1,088,274 text-image pairs and 1.17M annotated pairwise comparisons from state-of-the-art generative models together with low- and high-quality real-world images. Annotators choose between paired images for the same prompt using composite human preference, summarized in the paper as aesthetics, semantic similarity to the prompt, and overall coherence. This is not a classic scalar aesthetics dataset, but it is explicitly framed as the first wide-spectrum human preference dataset because it spans low-quality synthetic images through high-quality real photographs and covers multiple generator families and prompt categories (Ma et al., 5 Aug 2025).

6. Boundaries, controversies, and selection criteria

The field also contains datasets adjacent to aesthetics that widen a different variable than human beauty judgment. The Moonworks Lunara Aesthetic Dataset is a curated set of 2,000 image–prompt pairs with 17 region–style combinations, 7 topics, and structured annotations describing salient objects, attributes, relationships, and stylistic cues. Its aesthetic emphasis comes from curation and LAION Aesthetics v2 evaluation rather than per-image human ratings, and the paper reports a mean predicted score of 6.32 with 33.99% of images above 6.5; it is released under Apache 2.0 (Wang et al., 12 Jan 2026). Hyper-Skin broadens “aesthetics” in a different sense again: it is a 400–1000 nm facial skin hyperspectral dataset with 330 hyperspectral cubes in the abstract and 306 in the methods section, collected from 51 subjects, with paired hyperspectral and synthetic RGB data for skin-spectrum reconstruction. Its relevance is to skin appearance, cosmetology, and skin well-being rather than to generic subjective beauty labels (Ng et al., 2023).

These boundary cases clarify an important controversy. A large, style-rich, or appearance-oriented dataset is not automatically a human-judgment aesthetics dataset. Lunara is optimized for high-quality stylistic conditioning and legal reusability, but it does not report human aesthetic ratings per image (Wang et al., 12 Jan 2026). Hyper-Skin is wide-spectrum in wavelength coverage and calibrated skin reflectance, but not in subjective aesthetic judgments (Ng et al., 2023). Likewise, HPDv3 is explicitly a human preference dataset in which aesthetics is one component of a broader construct that also includes prompt alignment and coherence (Ma et al., 5 Aug 2025).

Across the field, several limitations recur. Some benchmarks omit exact split details or full preprocessing protocols; AVA-based work itself notes missing split counts and socially situated ratings from DPChallenge (Wang et al., 2016). HumanBeauty does not report a validation split or standard inter-annotator agreement coefficients (Liao et al., 31 Mar 2025). BAID derives scores from vote counts rather than direct per-image rating histograms and lacks detailed voter information (Yi et al., 2023). RAD gains scale by automatic generation but openly omits historical background and cultural significance (Liu et al., 29 Dec 2025). AesVideo-Bench does not report a standard numerical inter-annotator agreement statistic (Han et al., 30 Apr 2026). AesEval-Bench is graphic-design-specific and relies heavily on synthetic perturbations (An et al., 1 Mar 2026). The broader comparative analysis therefore argues that discrepancies between datasets “call into question the generalizability of previous research findings on single datasets” (Bartho et al., 2023).

A plausible implication is that there is no universally best wide-spectrum aesthetics dataset. AVA remains central when the goal is broad photographic judgments with multiple ratings and style metadata; MSC is preferable when beauty bias and semantic confounds must be minimized; BAID, RAD, and CompArt are better matches for artistic-image aesthetics; HumanBeauty targets human-centric imagery; ArtiMuse-10K is particularly suited to expert-grounded multimodal critique across several media; AesEval-Bench and AesVideo-Bench are appropriate when aesthetics must be grounded spatially or temporally; and HPDv3 is the strongest fit when prompt-conditioned human preference over real and generated images is the target construct. Taken together, these datasets show that “wide-spectrum aesthetics” is a plural concept: it may refer to broader value ranges, broader media coverage, broader annotation structure, or broader task scope, and the most informative benchmarks are those that make that breadth explicit rather than hiding it behind a single score.