BodyRealism
收藏资源简介:
BodyRealism是由亚马逊团队创建的一个专注于人体真实性的数据集,包含约30,000张生成的和真实的图像,这些图像与多模态信号(如文本描述、高质量的人体真实性评分和3D人体表示)配对。数据集的创建过程包括使用SOTA模型生成图像,并通过专家注释者进行人体真实性评分。BodyRealism旨在解决文本到图像生成模型中人体真实性评估的问题,特别是在生成图像中人体特征的真实性评估。
BodyRealism is a dataset focused on human body realism created by the Amazon team. It contains approximately 30,000 generated and real images paired with multimodal signals including text descriptions, high-quality human body realism scores, and 3D human body representations. The dataset construction process involves using SOTA models to generate images, followed by human body realism scoring by expert annotators. BodyRealism aims to address the problem of human body realism evaluation in text-to-image generation models, particularly the realism assessment of human body features in generated images.




