Human Style Preferences Images 图像生成偏好数据集

Human Style Preferences Images 数据集是一个用于评估文本到图像生成模型的人类标注数据集，由 Rapidata 利用 Rapidata 平台的创新性数据标注技术通过 Rapidata Python API 在短短 4 天内收集而成，并于 2025 年发布。该数据集通过展示两张图片并询问参与者哪张图片看起来更不奇怪或不自然，来收集人类对图像生成模型的一致性评价，包含了超过 120 万次的人类一致性投票，这些投票在不到 100 小时内完成，展示了 Rapidata 平台在数据收集速度上的优势。数据集具有大规模、全球代表性、多样化的提示和领先模型的比较等特点。该数据集对于基准测试新的图像生成模型、开发更好的生成模型评估指标、理解全球对 AI 生成图像的偏好、训练和微调图像生成模型以及研究跨文化审美偏好等方面具有重要价值。

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 using the innovative data annotation technology of the Rapidata platform via the Rapidata Python API, and released in 2025. This dataset collects human consensus evaluations of image generation models by displaying two images and asking participants which one appears less weird or unnatural. It contains over 1.2 million human consensus votes, which were completed in less than 100 hours, demonstrating the advantages of the Rapidata platform in terms of data collection speed. The dataset features large scale, global representativeness, diverse prompts, and comparisons of leading models. This dataset holds significant value for benchmarking new image generation models, developing better evaluation metrics for generative models, understanding global preferences for AI-generated images, training and fine-tuning image generation models, and researching cross-cultural aesthetic preferences.

搜集汇总

背景与挑战

背景概述

该数据集是一个用于评估文本到图像生成模型的人类标注数据集，通过收集人类对生成图像自然度的偏好投票构建而成。它规模庞大，具有全球代表性和多样化的提示，适用于模型基准测试、评估指标开发以及研究跨文化图像偏好。