3DGen-Bench
收藏资源简介:
3DGen-Bench是一个全面的3D模型生成人类偏好数据集,由上海人工智能实验室、浙江大学等机构共同构建。该数据集通过3DGen-Arena平台以对战方式高效收集人类偏好数据,包含500多个文本提示和500多个图像提示,覆盖270多个独特类别,涉及6个基本主题。数据集基于标注数据训练了两个评分模型,用于自动评估与人类判断相一致的质量评分。
3DGen-Bench is a comprehensive human preference dataset for 3D model generation, jointly constructed by institutions including Shanghai AI Laboratory and Zhejiang University. The dataset efficiently collects human preference data via the 3DGen-Arena platform in a battle-style competitive manner, containing over 500 text prompts and over 500 image prompts, covering more than 270 unique categories and involving 6 basic themes. Two scoring models were trained based on the annotated data from this dataset, which are used to automatically produce quality scores consistent with human judgments.




