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

Systematic and Comprehensive Categorization: We develop a taxonomy that systematically captures the core capabilities and application scenarios of T2I models.
Multiple Test Points per Prompt: To better reflect the user perspective, Magic-Bench-377 embeds multiple capabilities within a single prompt.
Clarity and Visualizability: Prompts should be concise, explicit and easy to visualize, while avoiding vague or non-visualizable descriptions
Neutrality and Fairness.: Descriptions involving regional specificity, references to celebrities, or copyrighted characters must be avoided to ensure fair evaluation across all models.

📊 Dataset Structure

Field	Description
`Prompt_text_cn`	Chinese version of the prompt
`Prompt_text_en`	English version of the prompt
`Application Scenario`	To account for the diversity of real world, magic bench 377 is divided into five categories
`Expression Form`	Refers to semantic units not directly pointing to visual elements but testing model’s understanding and reasoning over special forms of expressio
`Element Composition`	Refers to visual elements or information arising from the combination of multiple element
`Element`	Refers to visual elements or information that can be expressed by a single semantic unit, typically a word

🏷️ Taxonomy introduction

1. Application Scenario

Aesthetic design: Focuses on model use as a visual tool in professional design contexts, such as poster design, logo design, product design, etc. Models are expected to provide visually appealing outputs with high aesthetic quality.
Art : Focuses on user needs for high-level artistic creation, requiring models to generate outputs aligned with artistic styles, aesthetics, and visual imagination, such as oil painting, watercolor, sketching, or abstract expression.
Entertainment : Focuses on user needs for casual, creative and entertaining content, often reflecting internet culture (e.g., memes, emojis, or playful illustrations). The goal is to stimulate fun, amusement, or humor.
Film : Focuses on user needs for story-driven content creation, such as storyboards, cinematic scenes, or animated sequences. Models are expected to understand narrative details and generate scenes with coherent environments and character interactions.
Functional design : Focuses on user needs for practical work and learning materials, such as teaching slides, product manuals, or office diagrams. Outputs emphasize clarity, conciseness and informativeness.

2. Expression Form

Pronoun Reference: Pronouns (he, she, it, they) referring back to entities mentioned earlier in the text, requiring the model to resolve co-reference.
Negation: Negative expressions such as "no", "without" or "does not".
Consistency: Multiple entities of the same type sharing the same attribute.

3. Element Composition

Anti-Realism: Combinations that contradict real-world cognition or physical laws.
Multi-Entity Feature Matching: Multiple entities of the same type with distinct attribute values.
Layout & Typography: Descriptions of spatial or positional relationships among images, text, or symbols.

4. Element

Entity: Semantic units referring to entities such as people, animals, scenes, costumes, and decorations, including real-world and virtual entities, man-made objects, and natural elements.
Entity Description: Semantic units describing the quantity, attributes, forms, states, or relationships of entities.
Image Description: Semantic units that describe visual elements of a scene, including style, aesthetics, and artistic knowledge.

📁 Files

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

🚀 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')

📈 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
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