EmoArt: Emotion-Aware Artistic Datasets

• EmoArt is a suite of emotion-centered art datasets that integrate visual content, formal attributes, and affective labels for robust art synthesis and evaluation.
• The multidimensional dataset includes 132,664 paintings with detailed annotations on style, content, arousal–valence, and emotion categories, enriching both generative tasks and interpretability.
• The CoEmoGen subset and salience extension demonstrate scalable, caption-driven art generation and attribute-grounded selective reasoning to isolate emotionally operative features.

EmoArt is an emotion-centered designation in recent computational art, affective computing, and generative modeling literature for datasets and benchmarks that connect artistic imagery, formal visual properties, and affective supervision. The name is used most prominently in two ways: as a multidimensional painting dataset of 132,664 artworks with structured annotations for content, formal attributes, valence–arousal, emotion categories, and therapeutic potential (Zhang et al., 4 Jun 2025), and as a 13,633-image WikiArt-based artistic extension used to demonstrate the scalability of CoEmoGen for emotional image content generation (Yuan et al., 5 Aug 2025). A later line of work extends the former with a 1,400-artwork human-salience benchmark for attribute-grounded selective reasoning about which formal cues are emotionally operative in a given artwork (Zhang et al., 15 May 2026).

1. Terminology and referents

In current usage, EmoArt is not a single method. It denotes dataset-level resources that occupy different positions in the pipeline of emotion-aware artistic AI: corpus construction, supervised generation, and explanation-grounded reasoning. This naming overlap is consequential because the two principal EmoArt resources differ substantially in scale, annotation granularity, and intended task formulation (Zhang et al., 4 Jun 2025).

EmoArt usage	Scale	Primary role
Multidimensional EmoArt	132,664 artworks	Emotion-aware artistic generation and benchmarking
CoEmoGen EmoArt	13,633 art images	Scalable art-domain extension for EICG
EmoArt salience extension	1,400 artworks	Attribute-grounded selective reasoning

The larger multidimensional resource is explicitly framed as a response to the fact that many earlier emotion datasets are photo-centric, provide only coarse emotion labels, or lack the fine-grained visual attribute annotations needed for emotion-grounded artistic synthesis (Zhang et al., 4 Jun 2025). The CoEmoGen version instead serves as a proof-of-concept for scalable art-domain expansion using MLLM-generated sentence-level captions rather than manually curated word-level attributes (Yuan et al., 5 Aug 2025). The salience extension recasts EmoArt as a benchmark for explanation quality, asking not merely which attributes are visible, but which ones actually support the affective judgment for a specific artwork (Zhang et al., 15 May 2026).

2. Multidimensional dataset design

The multidimensional EmoArt dataset contains 132,664 artworks across 56 painting styles, drawn from open or public collections including WikiArt, The Met, Europeana, and other museum or open-access sources. The authors report an initial collection of over 200,000 artworks from more than 150 styles, followed by four filtering stages: art form filtering to keep only paintings, content safety filtering, image quality filtering to remove images below 300×300 or with artifacts, and category balance filtering to remove styles with fewer than 400 samples (Zhang et al., 4 Jun 2025).

Its central contribution is a five-part annotation schema intended to bind depicted content, formal visual structure, and affective interpretation into a unified supervisory signal. Each image includes:

• an objective scene/content description;
• five visual attributes: Brushwork, Composition, Color, Line, and Light;
• binary arousal–valence labels based on Russell’s circumplex model;
• 12 representative emotion categories;
• a label for potential art-therapy effects.

The arousal–valence layer is defined as

$$\text{Arousal} \in \{\text{High}, \text{Low}\}, \qquad \text{Valence} \in \{\text{Positive}, \text{Negative}\}.$$

The twelve emotion categories are distributed across the arousal–valence plane: aroused, excited, and happy for positive high arousal; alarmed, annoyed, and frustrated for negative high arousal; sad, bored, and tired for negative low arousal; and content, calm, and glad for positive low arousal (Zhang et al., 4 Jun 2025).

The descriptions are intended to exceed ordinary captioning in affective and formal specificity. The paper reports an average of 35.6 words per description, compared with 15.8 for ArtEmis and 10.5 for COCO captions. Annotation was produced through GPT-4o-assisted generation with human verification and multi-round checking, and validated on 5,600 images across the 56 styles by 10 trained annotators. Reported agreement with human labels is 98.01% for descriptions, 98.56% for visual attributes, and 91.47% for emotion; the paper also reports Gwet’s AC1 and McNemar’s test (Zhang et al., 4 Jun 2025).

The dataset is affectively imbalanced in a way that is itself analytically useful. The reported distribution is 87.93% Positive valence, 76.41% Low arousal, with Calm at 55.95%, Excited at 15.50%, and Contentment at 15.35%. The paper further notes style-dependent tendencies: Realism and Romanticism tend to be calm and peaceful; Expressionism and Surrealism show more high-arousal and negative affect; Chinese painting, Ukiyo-e, and Gongbi are strongly associated with low-arousal positive emotion; and Social Realism shows higher negative valence and alarm (Zhang et al., 4 Jun 2025).

3. Scalable art-domain extension for CoEmoGen

Within CoEmoGen, EmoArt is presented as the first large-scale emotional art image dataset and functions as a scalable artistic extension of an existing emotional image content generation pipeline. It contains 13,633 emotionally representative artistic images collected from WikiArt, spanning 129 artists, 11 genres, and 27 styles (Yuan et al., 5 Aug 2025).

Its construction pipeline is deliberately standardized. The authors collect about 100,000 WikiArt images, predict an emotional category for each image using a classifier trained on EmoSet, retain only those with emotion confidence above 0.75, and then generate one-sentence captions with the same MLLM-based captioning procedure used elsewhere in CoEmoGen. To reduce hallucination, they compute CLIP similarity for each image–caption pair and discard the bottom 20% of samples in each emotion category (Yuan et al., 5 Aug 2025).

A notable design decision is emotion-set reduction. The Mikels model originally supplies eight categories—amusement, awe, contentment, excitement, anger, disgust, fear, and sadness—but the paper states that excitement and disgust are rare in artistic expressions and account for less than 1% in this context. EmoArt therefore focuses on six emotions: amusement, awe, contentment, anger, fear, and sadness (Yuan et al., 5 Aug 2025).

The captioning protocol makes the emotion label explicit prior knowledge. The MLLM is prompted to produce a one-sentence description emphasizing emotion-related visual details such as brightness, colorfulness, scene type, object classes, facial expressions, and human actions. This makes EmoArt, in the CoEmoGen sense, not merely a curated image set but an instantiated example of a sentence-level semantic supervision pipeline meant to replace weaker word-level attribute labels (Yuan et al., 5 Aug 2025).

4. Role in emotion-aware artistic generation

The multidimensional EmoArt resource is explicitly designed to support generation as well as understanding. Its prompt template is formulated as

$$\text{Style} + \text{Arousal} + \text{Valence} + \text{Description},$$

thereby making emotion a primary conditioning signal rather than an auxiliary tag. For evaluation, the paper uses FID, SSIM, PSNR, LPIPS, and a proposed Attributes Alignment metric. The latter is computed by fine-tuning MiniCPM-V-2.6 on EmoArt and measuring similarity to the target text in CLIP embedding space. Benchmarking includes FLUX.1-dev, FLUX.1-schnell, SDXL, Stable Diffusion 3.5, PixArt-sigma, Playground, and OpenJourney; a LoRA fine-tuned FLUX.1-dev model trained on 50 paintings per artistic category and conditioned on Description, Arousal, and Valence performs best on most subjective and artistic measures, including brushstroke, color, composition, line quality, and overall quality (Zhang et al., 4 Jun 2025).

The CoEmoGen version of EmoArt serves a different generative purpose. There, EmoArt is not a separate model but a demonstration that sentence-level caption supervision can be transferred to a new artistic domain. The framework combines MLLM-generated emotion-triggering captions with HiLoRA, a parameter-efficient module based on the claim that emotions of the same polarity share low-level attributes but differ in high-level semantics. HiLoRA contains 2 polarity-shared LoRAs and 8 emotion-specific LoRAs, with an update rule

$W' = W + \Delta W = W + A \cdot B,$

and, for a positive-polarity example such as amusement,

$$W' = W + \Delta W_1^p + \Delta W_2^e = W + A_1^p B_1^p + A_2^e B_2^e.$$

In this interpretation, the polarity-shared component captures common low-level cues such as brightness and colorfulness, while the emotion-specific component captures finer semantics (Yuan et al., 5 Aug 2025).

The overall implication is that EmoArt-based generation research has split into two complementary strategies. One emphasizes rich multidimensional supervision and benchmarking across mainstream diffusion backbones; the other emphasizes scalable caption-centric expansion into new artistic domains. This suggests that “EmoArt” has become a focal term for moving emotional image generation beyond coarse class labels and toward art-specific conditioning structures.

5. Salience-grounded reasoning and explanation

The 2026 extension of EmoArt reframes the resource as a benchmark for artwork emotion understanding under Attribute-Grounded Selective Reasoning (AGSR). The key claim is that MLLMs often suffer from attribute flooding: they enumerate many visible formal properties without isolating which ones actually support the emotion judgment. To make this measurable, the authors add a 1,400-artwork human-salience extension annotated by 15 annotators with formal art-training backgrounds (Zhang et al., 15 May 2026).

The formal attribute vocabulary is

$$\mathcal{A}=\{\mathrm{color},\mathrm{composition},\mathrm{line},\mathrm{light},\mathrm{brushstroke}\}.$$

For each artwork, annotators assign binary salience decisions over these five attributes. The central distinction is between attributes that are merely present and attributes that are emotionally operative. The modeling desideratum is sparse evidence selection:

$\|\mathbf{s}(x)\|_0 \ll |\mathcal{A}|.$

The proposed method, FAB-G (Formal-Attribute Bottleneck-Guided reasoning), operationalizes AGSR in two stages. First, five attribute-specific agents predict salience decisions,

$s_a = g_a(x;\theta_a) \in \{0,1\},$

yielding the retained set

$\mathcal{S}(x)=\{a \in \mathcal{A} \mid s_a=1\}.$

Second, a final analysis agent generates affective predictions and explanations conditioned on the selected subset rather than the full attribute inventory (Zhang et al., 15 May 2026).

On Qwen3-VL, the reported results are: Base Model at 29.33% emotion, 56.67% arousal, and 80.00% valence; CoT-SFT at 48.67%, 78.00%, and 88.00%; and FAB-G at 50.00%, 82.00%, and 90.00%. Evidence alignment is measured by Dice and Tversky; the best FAB-G scores on EmoArt are 0.8742 / 0.8808 sample-wise mean and 0.8450 / 0.8454 attribute-wise mean, with $\alpha=0.8$ and $\beta=0.2$ in the Tversky metric to penalize false positives more heavily. Explanations also become markedly shorter: for Qwen3-VL, FAB-G produces 57.56 tokens, compared with 145.71 for CoT and 140.15 for One-shot prompting. Cross-dataset evaluation on Abstract, ArtEmis, and WikiArt further suggests transfer beyond the source distribution, while exposing that Line, Light, and Brushstroke remain harder than Color and Composition (Zhang et al., 15 May 2026).

6. Methodological issues, limitations, and significance

Several methodological constraints recur across the EmoArt literature. First, the multidimensional EmoArt dataset is limited to paintings, not all art media, and removes styles with fewer than 400 samples, improving balance at the cost of reducing long-tail coverage (Zhang et al., 4 Jun 2025). Second, the CoEmoGen EmoArt subset depends on classifier-based emotion filtering, MLLM caption quality, and CLIP-based filtering, so its quality is partly inherited from upstream components (Yuan et al., 5 Aug 2025). Third, even the salience extension models only five formal attributes, while emotional interpretation may also depend on iconography, cultural background, and viewer experience; the salience subset is also modest relative to the full corpus, and Light is reported as especially difficult to transfer reliably across datasets (Zhang et al., 15 May 2026).

A common misconception is to treat art emotion as a single-label property. Earlier work already showed that affective interpretation in art is substantially subjective: ArtEmis reports 439,121 emotion attributions and explanations on 81,446 artworks, and 61% of artworks receive at least one positive and one negative emotion across annotators (Achlioptas et al., 2021). EmoArt’s design choices can be read as responses to that fact. The multidimensional version introduces layered supervision rather than a single target label; the CoEmoGen version reduces the label set where artistic frequency is low; and the salience extension separates visible attributes from emotionally admissible evidence.

Within the broader research trajectory, EmoArt sits between affective language resources and emotion-conditioned generation frameworks. ArtEmis contributes grounded human explanations for why artworks evoke emotions, but not the five-attribute structured supervision used in EmoArt (Achlioptas et al., 2021). EmoGen formulates Emotional Image Content Generation (EICG) and introduces evaluation concepts such as Emo-A, Sem-C, and Sem-D for emotion-faithful generation (Yang et al., 2024). EmoArt extends this agenda into art-specific corpora where content, style, formal properties, and affective labels are jointly organized.

Taken together, the EmoArt literature defines a research program rather than a single artifact: building art-domain resources in which what is depicted, how it is formally rendered, and what it is felt to express become computationally linked. That program now spans dataset curation, multimodal caption induction, diffusion benchmarking, and bottlenecked reasoning about emotionally operative evidence.