# MagicBrush: A Manually Annotated Dataset for Instruction-Guided Image Editing

**Authors:** Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, Yu Su  
**arXiv:** [2306.10012](https://arxiv.org/abs/2306.10012)  
**Date:** 2023-06-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2306.10012v1/

## Key points

- MagicBrush is a new dataset for instruction-guided real image editing
- It addresses the limitations of existing methods for text-guided image editing in daily life and professional applications such as Photoshop
- Current approaches are either zero-shot or trained on an automatically synthesized dataset with high noise levels, requiring significant manual tuning to achieve desirable outcomes in practice
- MagicBrush comprises over 10,000 manually annotated triples covering diverse scenarios such as single-turn, multi-turn, mask-provided, and mask-free editing
- This dataset supports training large-scale text-guided image editing models and fine-tuning InstructPix2Pix on MagicBrush produces much better images according to human evaluation
- The authors conducted extensive experiments to evaluate current image editing baselines from multiple dimensions including quantitative, qualitative, and human evaluations
- The results reveal the challenging nature of the MagicBrush dataset and highlight the gap between current baselines and real-world editing needs
- Overall, MagicBrush provides a valuable resource for advancing text-guided image editing techniques towards more practical applications in various domains.

## Summary

MagicBrush is a new, large-scale dataset for instruction-guided real image editing that addresses the limitations of existing methods. Text-guided image editing is essential in daily life and professional applications such as Photoshop, but current approaches are either zero-shot or trained on an automatically synthesized dataset with high noise levels. As a result, these methods require significant manual tuning to achieve desirable outcomes in practice. MagicBrush comprises over 10,000 manually annotated triples (source image, instruction, target image) covering diverse scenarios such as single-turn, multi-turn, mask-provided, and mask-free editing. This dataset supports training large-scale text-guided image editing models and fine-tuning InstructPix2Pix on MagicBrush produces much better images according to human evaluation. The authors further conduct extensive experiments to evaluate current image editing baselines from multiple dimensions including quantitative, qualitative, and human evaluations. The results reveal the challenging nature of the MagicBrush dataset and highlight the gap between current baselines and real-world editing needs. Overall, MagicBrush provides a valuable resource for advancing text-guided image editing techniques towards more practical applications in various domains.