SlowQZK123/EmoArt-130k

# EmoArt: A Large-Scale Emotion-Annotated Artistic Dataset

[](https://huggingface.co/datasets/printblue/EmoArt-130k)
[](LICENSE)

## Overview

**EmoArt** is a comprehensive, large-scale emotion-annotated artistic dataset containing **132,664 high-resolution artworks** spanning **56 painting styles** across **7 thematic categories**. This dataset bridges the gap between visual art and emotional computing, enabling groundbreaking research in emotion-aware AI systems.

### Key Statistics
- 📊 **132,664 artworks** with rich emotional annotations
- 🎨 **56 distinct painting styles** from global art traditions
- 🌍 **7 thematic categories** covering diverse artistic movements
- 🧠 **Multi-dimensional emotion labeling** with therapeutic insights
- 🤖 **GPT-4o enhanced annotations** with human validation

## Dataset Versions

We provide two carefully curated versions to meet different research needs:

| Version | Size | Description | Use Case |
|---------|------|-------------|----------|
| **[EmoArt-130k](https://huggingface.co/datasets/printblue/EmoArt-130k)** | 132,664 images | Complete dataset with full coverage | Production models, comprehensive research |
| **[EmoArt-5k](https://huggingface.co/datasets/printblue/EmoArt-5k)** | 5,600 images | Curated subset (100 per style) | Prototyping, education, resource-constrained environments |

## Quick Start

### Download by Category

Access individual artistic categories as needed:

```bash
# Download specific category (e.g., Cubism)
wget https://huggingface.co/datasets/printblue/EmoArt-130k/resolve/main/Cubism.tar.gz
tar -xzvf Cubism.tar.gz

# Download all annotations (single file)
wget https://huggingface.co/datasets/printblue/EmoArt-130k/resolve/main/Annotation.json
```

### Load with Python

```python
from datasets import load_dataset

# Authenticate with Hugging Face
# Run: huggingface-cli login

# Load the complete dataset
ds = load_dataset("printblue/EmoArt-130k")
```

## Dataset Architecture

### File Organization

The dataset follows a modular structure for efficient access and storage:

```
EmoArt Dataset/
├── Classics.tar.gz # Traditional art styles
├── Modern_Edge.tar.gz # Modern movements
├── East_Spirit.tar.gz # Asian art traditions
├── Chromatic_Soul.tar.gz # Color-focused styles
├── Dream_Visions.tar.gz # Surreal and abstract works
├── Form_Flow.tar.gz # Minimalist and geometric styles
├── Social_Mirror.tar.gz # Social and political art
└── Annotation.json # Unified annotation file
```

**Architecture Benefits:**
- 🔄 **Modular downloads**: Access only needed categories
- 📁 **Centralized annotations**: Single JSON file for all metadata
- 💾 **Storage efficient**: Compressed tar.gz format
- 🚀 **Fast experimentation**: Category-specific access

### Individual Category Structure

Each category archive contains organized image files:

```
{Category}.tar.gz
└── images/
├── image_00001.jpg
├── image_00002.jpg
├── image_00003.jpg
└── ... (category-specific count)
```

## Annotation Framework

### Comprehensive Annotation Schema

The unified `Annotation.json` provides rich metadata for every artwork:

```json
{
"request_id": "{Art_Category}_request-1",
"description": {
"first_section": {
"description": "Detailed visual and compositional analysis of the artwork"
},
"second_section": {
"visual_attributes": {
"brushstroke": "Technical analysis of brushwork and application",
"color": "Color palette, saturation, and harmony assessment",
"composition": "Structural organization and visual flow analysis",
"light_and_shadow": "Illumination, contrast, and depth evaluation",
"line_quality": "Line characteristics and stylistic execution"
},
"emotional_impact": "Comprehensive emotional response and psychological effect analysis"
},
"third_section": {
"emotional_arousal_level": "High/Low",
"emotional_valence": "Positive/Negative",
"dominant_emotion": "Primary emotional category",
"healing_effects": ["Therapeutic and wellness applications"]
}
},
"image_path": "Images\\{Category}\\{filename}.jpg"
}
```

### Annotation Dimensions

Each artwork includes multi-layered emotional and visual analysis:

#### Emotional Intelligence
- **12 Core Emotions**: Joy, Sadness, Anger, Fear, Surprise, Disgust, Calm, Excitement, and more
- **Valence Classification**: Positive/Negative emotional tone
- **Arousal Assessment**: High/Low emotional intensity
- **Dominant Response**: Primary emotional impact

#### Visual Analysis Framework
- **Brushstroke Technique**: Application method and artistic execution
- **Color Psychology**: Palette choices and emotional resonance
- **Compositional Structure**: Visual organization and flow
- **Light Dynamics**: Illumination patterns and contrast usage
- **Linear Expression**: Line quality and stylistic character

#### Therapeutic Applications
- **Healing Potential**: Stress relief, mood enhancement, contemplative benefits
- **Wellness Integration**: Applications in art therapy and mental health

## Artistic Categories

### Thematic Organization

The dataset encompasses seven major artistic domains:

| Category | Focus | Key Styles | Cultural Scope |
|----------|-------|------------|----------------|
| **Classics** | Traditional mastery | Realism, Renaissance, Baroque, Neoclassicism | Western classical tradition |
| **Modern Edge** | Revolutionary movements | Cubism, Expressionism, Futurism, Dadaism | Early 20th century innovation |
| **East Spirit** | Asian traditions | Chinese Painting, Ukiyo-e, Sumi-e, Miniatures | East Asian artistic heritage |
| **Chromatic Soul** | Color exploration | Fauvism, Color Field, Abstract Expressionism | Color-centric movements |
| **Dream Visions** | Surreal imagination | Surrealism, Abstract Art, Symbolism | Subconscious and dreamlike art |
| **Form & Flow** | Geometric precision | Minimalism, Constructivism, Op Art | Structural and mathematical art |
| **Social Mirror** | Cultural commentary | Social Realism, Political Art, Street Art | Socially engaged art |

### Style Coverage

The dataset represents the full spectrum of global artistic expression across cultures and time periods, ensuring comprehensive coverage for cross-cultural emotion research.

## Performance Benchmarks

### Emotion-Aware Generation Models

Evaluation results on state-of-the-art diffusion models:

| Model | Image Quality ↑ | Emotion Alignment ↑ | FID Score ↓ | Training Efficiency |
|-------|----------------|-------------------|-------------|-------------------|
| **FLUX.1-dev-lora** | **0.6604** | **0.6698** | 31.65 | High |
| PixArt-sigma | 0.6505 | 0.6342 | 36.23 | Medium |
| FLUX.1-dev | 0.6392 | 0.6228 | 21.29 | Medium |
| Playground | 0.6486 | 0.6247 | 42.57 | Low |

*Higher scores indicate better performance for Quality and Emotion Alignment; lower FID scores indicate better image quality.*

## Research Applications

### Core Research Areas
- **Emotion-Aware AI**: Training models that understand and generate emotionally resonant content
- **Affective Computing**: Bridging human emotion and computational understanding
- **Cross-Modal Learning**: Vision-language models with emotional intelligence
- **Computational Aesthetics**: Quantifying beauty and emotional impact in art
- **Digital Art Therapy**: AI-assisted therapeutic applications

### Practical Applications
- **Content Generation**: Emotion-driven artistic creation
- **Therapeutic Tools**: AI-powered art therapy systems
- **Cultural Studies**: Cross-cultural emotion perception research
- **Educational Technology**: Interactive art history and emotion learning
- **Creative Industries**: Emotion-aware design and marketing tools

## Quality Assurance

### Ethical Standards
- ✅ **Open Access Only**: Exclusively public domain and Creative Commons works
- 🔍 **Content Filtering**: Manual review for sensitive or inappropriate material
- 🌍 **Cultural Balance**: Representative sampling across global art traditions
- 👥 **Human Oversight**: Expert validation of AI-generated annotations

### Technical Quality
- 🖼️ **High Resolution**: Professional-quality image standards
- 🤖 **AI-Human Hybrid**: GPT-4o annotations with human expert validation
- 📊 **Consistency Checks**: Standardized annotation protocols
- 🔄 **Continuous Improvement**: Community feedback integration

## Getting Started

### For Researchers
1. **Browse the dataset** on Hugging Face to understand scope and structure
2. **Start with EmoArt-5k** for initial experiments and prototyping
3. **Download specific categories** relevant to your research focus
4. **Scale to EmoArt-130k** for comprehensive model training

### For Developers
1. **Use the Python API** for seamless integration
2. **Implement modular loading** to manage memory efficiently
3. **Leverage the annotation structure** for multi-task learning
4. **Contribute improvements** back to the community

## Resources & Support

### Documentation & Code
- 📖 **Research Paper**: [MM'25 Conference Publication](https://arxiv.org/abs/2025.emoart)
- 💻 **Source Code**: [GitHub Repository](https://github.com/ZHILIANGZHANG/EmoArt-130k)
- 📊 **Benchmarks**: Performance baselines and evaluation metrics

### Community & Support
- 📧 **Contact**: [[e-mail]](mailto:[e-mail])
- 🐛 **Bug Reports**: [GitHub Issues](https://github.com/ZHILIANGZHANG/EmoArt-130k/issues)
- 💬 **Discussions**: [Hugging Face Community](https://huggingface.co/datasets/printblue/EmoArt-130k/discussions)

---

> *"Art enables us to find ourselves and lose ourselves at the same time."*
> **EmoArt enables AI to do the same.**