# HQ-Edit: A High-Quality Dataset for Instruction-based Image Editing

**Authors:** Mude Hui, Siwei Yang, Bingchen Zhao, Yichun Shi, Heng Wang, Peng Wang, Yuyin Zhou, Cihang Xie  
**arXiv:** [2404.09990](https://arxiv.org/abs/2404.09990)  
**Date:** 2024-04-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2404.09990v1/

## Key points

- Study introduces HQ-Edit dataset with approximately 200,000 edits
- Researchers developed scalable data collection pipeline using advanced foundation models
- Dataset quality ensured through diverse online examples, high-quality diptychs, and precise alignment techniques
- Proposed evaluation metrics - Alignment and Coherence - to assess image edit pair quality quantitatively
- HQ-Edit dataset features high-resolution images rich in detail and comprehensive editing prompts
- Enhances capabilities of existing image editing models, surpassing state-of-the-art performance when used with finetuned InstructPix2Pix model
- Project page for HQ-Edit: https://thefllood.github.io/HQEdit_web

## Summary

The study by Mude Hui, Siwei Yang, Bingchen Zhao, Yichun Shi, Heng Wang, Peng Wang, Yuyin Zhou, and Cihang Xie introduces , a high-quality  containing approximately 200,000 edits. Unlike previous methods that relied on attribute guidance or human feedback to construct datasets, the researchers developed a scalable data collection pipeline utilizing advanced foundation models such as  and . To ensure the dataset's high quality, a diverse range of examples were initially gathered online and then expanded upon. These examples were used to create high-quality diptychs showcasing input and output images alongside detailed text prompts. Precise alignment was achieved through post-processing techniques. Additionally, the authors proposed two evaluation metrics - Alignment and Coherence - to quantitatively assess the quality of image edit pairs using . features high-resolution images rich in detail and accompanied by comprehensive editing prompts. This dataset significantly enhances the capabilities of existing image editing models. For instance, when an HQ-Edit finetuned InstructPix2Pix model is utilized, it can achieve state-of-the-art image editing performance surpassing even those models fine-tuned with human-annotated data. The project page for HQ-Edit can be accessed at https://thefllood.github.io/HQEdit_web. This dataset represents a valuable resource for advancing research in instruction-based image editing and improving the performance of image editing models.