Emotion-Aware
Artistic Generation

Bridging the gap between visual art and emotion understanding with 132,000+ annotated artworks and multidimensional analysis.

EmoArt Overview Visualization
132K+ Artworks
56 Art Styles
12 Emotion Classes
5 Visual Attributes

Why EmoArt?

A pioneering large-scale dataset capable of powering the next generation of affective computing and creative AI.

Comprehensive Collection

Spanning Western and Eastern traditions, from Renaissance masterpieces to contemporary abstract movements.

GPT-4o Annotations

Rich, multi-dimensional annotations generated by advanced AI and validated by humans with >85% agreement.

Emotional Depth

Covering 12 representative emotions across the full valence-arousal spectrum for nuanced understanding.

Visual Analysis

Structured decomposition of Brushwork, Composition, Color, Line, and Light.

FAB-G Salience Extension

A compact human-validated extension for attribute-grounded artwork emotion reasoning.

Attribute Salience

EmoArt-Salience annotates which formal attributes are emotionally operative for each artwork, separating cues that are merely visible from cues that support the affective judgment.

Five Formal Cues

The salience labels cover Color, Composition, Line, Light, and Brushstroke, matching the FAB-G attribute-agent vocabulary used for selective reasoning.

FAB-G Pipeline

FAB-G first predicts salient attributes with specialized agents, then sends only the selected attributes to the final agent for cue-constrained emotion analysis.

Benchmark Performance

Evaluating state-of-the-art text-to-image diffusion models.

Model Overall Quality Emotion Alignment Color Composition FID ↓
FLUX.1-dev-lora 0.6604 0.6698 0.6974 0.6698 31.65
FLUX.1-dev 0.6392 0.6450 0.6503 0.6450 21.29
PixArt-sigma 0.6505 0.6342 0.6746 0.6342 36.64
Playground 0.6486 0.6247 0.6788 0.6247 42.57
SD3.5 0.4350 0.6324 0.3420 0.4324 37.96
Comparison

Visual Analysis

  • Color Field Painting: FLUX.1-dev-finetuned uses pure blue and white blocks to create a calming atmosphere, capturing the essence of Color Field Painting perfectly.
  • Traditional Chinese Painting: The fine-tuned model employs minimalist compositions and soft brushwork, conveying the serenity typical of East Asian aesthetics.
  • High-Arousal Emotions: For intense emotions, the model effectively utilizes chaotic line work and asymmetric compositions.

Dataset Composition

A deep dive into the artistic styles and emotional distributions.

Detailed Statistics by Style

Art Style Entries Arousal Distribution Valence Distribution Dominant Emotions
Abstract Art 1,759 High: 51.5% / Low: 48.5% Pos: 93.5% / Neg: 6.5% Excited (47%), Calm (39%)
Abstract Expressionism 3,674 High: 58.7% / Low: 41.3% Pos: 85.7% / Neg: 14.3% Excited (47%), Calm (31%)
Baroque 7,995 High: 21.4% / Low: 78.6% Pos: 86.2% / Neg: 13.8% Calm (55%), Contentment (19%)
Impressionism 11,736 High: 9.9% / Low: 90.1% Pos: 95.4% / Neg: 4.6% Calm (67%), Contentment (20%)
Realism 15,307 High: 8.0% / Low: 92.0% Pos: 88.9% / Neg: 11.1% Calm (65%), Contentment (20%)
Expressionism 10,065 High: 36.2% / Low: 63.8% Pos: 75.9% / Neg: 24.1% Calm (42%), Excited (21%), Sad (11%)
Chinese Painting 4,157 High: 0.2% / Low: 99.8% Pos: 99.9% / Neg: 0.1% Calm (89%), Contentment (10%)
Ukiyo-e 1,730 High: 13.9% / Low: 86.1% Pos: 95.2% / Neg: 4.8% Calm (74%), Contentment (11%)
Full statistical breakdown available in the dataset documentation.

Citation

If you use EmoArt in your research, please cite our paper.

@inproceedings{zhang2025emoart, title={EmoArt: A Multidimensional Dataset for Emotion-Aware Artistic Generation}, author={Zhang, Cheng and Xie, Hongxia and Wen, Bin and Zuo, Songhan and Zhang, Ruoxuan and Cheng, Wen-Huang}, booktitle={Under Review}, pages={1--6}, year={2025} }