Magic Bench: A Comprehensive Text-to-Image Generation Evaluation Dataset

📖 Overview

Magic Bench is a comprehensive evaluation dataset designed for text-to-image generation models. It contains 377 carefully curated prompts with detailed annotations across multiple dimensions, providing both Chinese and English versions for cross-lingual evaluation.

🎯 Dataset Features

377 evaluation prompts covering diverse scenarios
Bilingual support: Both Chinese and English prompts
Multi-dimensional annotations: 9 different evaluation dimensions
Comprehensive coverage: Aesthetic design and artistic photography scenarios

📊 Dataset Structure

The dataset includes the following fields:

Field	Description
`prompt_text_cn`	Chinese version of the prompt
`prompt_text_en`	English version of the prompt
`应用场景` (Application Scenario)	The application context
`表达形式` (Expression Form)	Form of expression annotations
`要素组合` (Element Combination)	Element combination patterns
`实体描述` (Entity Description)	Entity description types
`是否有风格` (Has Style)	Whether the prompt includes style specifications
`是否有美学知识` (Has Aesthetic Knowledge)	Whether aesthetic knowledge is required
`是否有氛围` (Has Atmosphere)	Whether atmospheric elements are present

🏷️ Annotation Dimensions

1. Application Scenario (应用场景)

创意设计 (Aesthetic design): Logo design, character design, product design, etc.
艺术 (art): Photography, artistic creation, etc.
个性化娱乐 (entertainment): Entertainment and personalized content
影视与故事 (film): Film and storytelling scenarios
效率提升 (functional design): Efficiency and functional design

2. Expression Form (表达形式)

无 (None): No specific form requirements
代词指代 (Pronoun Reference): Contains pronoun references
否定 (Negation): Contains negative expressions
统一性 (consistency): Requires consistent elements

3. Element Combination (要素组合)

无 (None): Single element
反现实 (Anti-Realism): Anti-realistic combinations
多实体多要素 (Multi-Entity Feature Matching): Complex multi-entity combinations
布局与排版 (Layout & Typography): Specific layout requirements

4. Entity Description (实体描述)

无 (None): No specific entity descriptions
属性 (attribute): Attribute descriptions
关系 (relation): Relationship descriptions
动作状态 (Action/State): Action or state descriptions
数量 (quantity): Quantity specifications

5. Style Specification (是否有风格)

是 (Yes): Contains specific style requirements
否 (No): No style specifications

6. Aesthetic Knowledge (是否有美学知识)

是 (Yes): Requires aesthetic understanding
否 (No): No aesthetic knowledge required

7. Atmospheric Elements (是否有氛围)

是 (Yes): Contains atmospheric descriptions
否 (No): No atmospheric elements

📁 Files

magic_bench_dataset.csv: Complete dataset with all annotations
magic_bench_chinese.csv: Chinese prompts with annotations
magic_bench_english.csv: English prompts with annotations

🚀 Usage

import pandas as pd

# Load the complete dataset
df = pd.read_csv('magic_bench_dataset.csv')

# Load Chinese version
df_cn = pd.read_csv('magic_bench_chinese.csv')

# Load English version
df_en = pd.read_csv('magic_bench_english.csv')

# Example: Filter prompts with style requirements
stylized_prompts = df[df['是否有风格'] == '是']

# Example: Get aesthetic design prompts
aesthetic_prompts = df[df['应用场景'] == '创意设计']

📈 Statistics

Total prompts: 377
Aesthetic design prompts: 95 (25.2%)
Art prompts: 80 (21.2%)
Prompts with style specifications: 241 (63.9%)
Prompts requiring aesthetic knowledge: 131 (34.7%)
Prompts with atmospheric elements: 22 (5.8%)

🎯 Use Cases

Model Evaluation: Comprehensive evaluation of text-to-image models
Benchmark Comparison: Compare different models across various dimensions
Research: Study model capabilities in different scenarios
Fine-tuning: Use as training or validation data for model improvement

📄 Citation

If you use this dataset in your research, please cite:

@dataset{magic_bench_377,
  title={Magic Bench: A Comprehensive Text-to-Image Generation Evaluation Dataset},
  author={outongtong},
  year={2025},
  email={outongtong.ott@bytedance.com},
  url={https://huggingface.co/datasets/ByteDance-Seed/MagicBench}
}

📜 License

This dataset is released under the cc-by-nc-4.0.

🤝 Contributing

We welcome contributions to improve the dataset. Please feel free to:

Report issues or suggest improvements
Submit pull requests with enhancements
Share your evaluation results using this dataset

📞 Contact

For questions or collaborations, please contact: outongtong.ott@bytedance.com

Keywords: text-to-image, evaluation, benchmark, dataset, computer vision, AI, machine learning