pamela
收藏资源简介:
PAMELA(Personalizing Text-to-Image Generation to Individual Taste)是一个专门为个性化研究构建的数据集,其核心内容是AI生成图像及其对应的人类美学质量评级。该数据集旨在支持个性化美学预测、视觉偏好的人口统计差异研究以及生成内容的奖励建模。数据集包含总计69,904个美学评级,覆盖5,077张独特的AI生成图像。这些图像分为‘艺术’和‘超写实’两种主要类型,并进一步细分为21个视觉主题组(如抽象、动物、建筑等),其中‘艺术’类图像应用了26种不同的艺术史风格(如印象派、立体主义等)。数据来自199名参与者,他们提供了年龄(19-60岁)、性别(男、女、不愿透露)和国籍(23个国家)等人口统计信息。每个数据条目都结构化为JSON格式,包含图像路径与ID、详细的图像元数据(组、风格、类型、生成提示)、匿名的参与者ID及其人口统计信息,以及美学评级(分为‘差’到‘优秀’五个等级的质量标签和1.0-5.0的连续原始分数)。数据集被精心划分为训练集(pamela_train,50,222个评级)、验证集(包含未见用户pamela_val_unseen和已知用户pamela_val_seen)和测试集(包含未见用户pamela_test_unseen和已知用户pamela_test_seen),以分别评估模型对已知用户新图像的预测能力以及对全新用户的泛化能力。数据集通过Krippendorffs α和组内相关系数等指标评估了评级者间的一致性。需要注意的是,参与者样本并非全球人口的代表,且某些人口统计子群体样本量较小。数据集在CC-BY-4.0许可下发布,图像由Nano Banana和FLUX.2模型生成。其预期用途明确为个性化美学研究,不适用于建立普适的美学标准或进行技术性的图像质量评估。
PAMELA (Personalizing Text-to-Image Generation to Individual Taste) is a dataset specifically constructed for personalized research, whose core content consists of AI-generated images and their corresponding human aesthetic quality ratings. This dataset is designed to support personalized aesthetic prediction, research on demographic differences in visual preferences, and reward modeling for generated content. It contains a total of 69,904 aesthetic ratings covering 5,077 unique AI-generated images. These images are divided into two primary categories: 'Art' and 'Photorealistic', and further subdivided into 21 visual theme groups (e.g., abstract, animals, architecture, etc.). Images in the 'Art' category adopt 26 distinct art historical styles such as Impressionism and Cubism. The dataset is sourced from 199 participants, who provided demographic information including age (19–60 years), gender (male, female, prefer not to say), and nationality from 23 countries. Each data entry is structured in JSON format, containing image path and ID, detailed image metadata (group, style, category, generation prompt), anonymized participant ID and their demographic details, as well as aesthetic ratings, which include quality labels ranging from 'Poor' to 'Excellent' and continuous raw scores from 1.0 to 5.0. The dataset is rigorously split into three subsets: the training set (pamela_train, 50,222 ratings), the validation set (including unseen-user pamela_val_unseen and seen-user pamela_val_seen), and the test set (including unseen-user pamela_test_unseen and seen-user pamela_test_seen), to respectively evaluate the model's predictive capability for new images of known users and its generalization ability to entirely new users. Inter-rater reliability of the dataset is assessed using metrics including Krippendorff's α and intraclass correlation coefficient. It is important to note that the participant sample is not representative of the global population, and some demographic subgroups have small sample sizes. The dataset is released under the CC-BY-4.0 license, and the images were generated using the Nano Banana and FLUX.2 models. Its intended use is explicitly limited to personalized aesthetic research, and it is not suitable for establishing universal aesthetic standards or conducting technical image quality assessment.




