HQ-Edit
收藏资源简介:
HQ-Edit是一个高质量的基于指令的图像编辑数据集,由加州大学圣克鲁兹分校创建,包含约200,000个编辑指令。该数据集通过利用GPT-4V和DALL-E 3等先进基础模型,构建了一个可扩展的数据收集管道。数据集的创建过程包括从在线资源收集多样化的示例,扩展这些示例,并创建包含详细文本提示的输入和输出图像的高质量双联画,随后通过后处理确保精确对齐。此外,HQ-Edit还提出了两个评估指标,即对齐度和一致性,以量化评估图像编辑对的质量。HQ-Edit的高分辨率图像和丰富的编辑提示显著增强了现有图像编辑模型的能力,特别是在解决图像编辑中的精确指令遵循问题。
HQ-Edit is a high-quality instruction-based image editing dataset created by the University of California, Santa Cruz, containing approximately 200,000 editing instructions. This dataset establishes a scalable data collection pipeline by leveraging advanced foundation models including GPT-4V and DALL-E 3. The dataset construction process involves collecting diverse samples from online resources, expanding these samples, generating high-quality input-output image pairs paired with detailed text prompts, and conducting post-processing to ensure precise alignment. Furthermore, HQ-Edit proposes two evaluation metrics, alignment and consistency, for quantitatively evaluating the quality of image editing pairs. The high-resolution images and abundant editing prompts of HQ-Edit significantly enhance the performance of existing image editing models, especially in resolving the issue of accurate instruction adherence in image editing.




