HATIE
收藏资源简介:
HATIE是一个大规模的文本引导图像编辑基准数据集,旨在解决当前图像编辑模型评估缺乏标准化方法的难题。该数据集基于GQA数据集,包含18,226张图像和19,933个可编辑对象,覆盖76个COCO类别。HATIE提供了丰富的编辑查询,包括对象添加、移除、替换、属性更改和背景风格变化等,并采用自动化和多方面的评估指标,如图像质量、对象保真度、背景保真度、对象一致性和背景一致性等,以与人类感知相一致。HATIE的创建为文本引导图像编辑领域的研究提供了可靠、客观和易于复现的评估方法。
HATIE is a large-scale text-guided image editing benchmark dataset developed to address the pressing challenge of lacking standardized evaluation methodologies for contemporary image editing models. Built upon the GQA dataset, HATIE comprises 18,226 images and 19,933 editable objects, covering 76 COCO categories. HATIE offers a wide range of editing queries, including object addition, removal, replacement, attribute modification, background style alteration, and more. It employs automated and multi-faceted evaluation metrics such as image quality, object fidelity, background fidelity, object consistency, and background consistency to align with human perceptual preferences. The creation of HATIE provides a reliable, objective, and readily reproducible evaluation framework for research in the field of text-guided image editing.
HATIE数据集概述
基本信息
- 数据集名称: HATIE (Human-aligned Benchmark for Text-guided Image Editing)
- 发布会议: CVPR 25 Highlight
- GitHub地址: https://github.com/SuhoRyu/HATIE
- arXiv论文: https://arxiv.org/abs/2505.00502
研究团队
- 主要作者:
- Suho Ryu
- Kihyun Kim
- Eugene Baek
- Dongsoo Shin
- Joonseok Lee (VIPLab, Seoul National University)
数据集特点
- 研究领域: 文本引导的图像编辑
- 目标: 建立可扩展的人类对齐基准
当前状态
- 代码和数据集: 即将发布




