Human Style Preferences Images Image Generation Preference Dataset

The Human Style Preferences Images dataset is a human-annotated dataset for evaluating text-to-image generation models. It was collected by Rapidata in just 4 days through the Rapidata Python API using the Rapidata platform's innovative data annotation technology and released in 2025. The dataset collects human consistency evaluations of image generation models by showing two pictures and asking participants which picture looks less strange or unnatural. It contains more than 1.2 million human consistency votes, which were completed in less than 100 hours, demonstrating the Rapidata platform's advantage in data collection speed. The dataset has the characteristics of large-scale, global representativeness, diverse prompts, and comparisons of leading models. The dataset is of great value for benchmarking new image generation models, developing better evaluation metrics for generation models, understanding global preferences for AI-generated images, training and fine-tuning image generation models, and studying cross-cultural aesthetic preferences.

The construction of the dataset not only relies on large-scale human voting, but also covers diverse participants from all over the world, ensuring the geographical and cultural representation of the data. In addition, carefully designed prompts are used in the construction of the dataset to test different aspects of the image generation model, thus ensuring the comprehensiveness and depth of the dataset.

This dataset is suitable for a variety of application scenarios, including but not limited to benchmarking new image generation models, developing evaluation metrics for generation models, understanding preferences for AI-generated images worldwide, training and fine-tuning image generation models, and studying cross-cultural aesthetic preferences.