# DragDiffusion: Harnessing Diffusion Models for Interactive Point-based Image Editing

**Authors:** Yujun Shi, Chuhui Xue, Jiachun Pan, Wenqing Zhang, Vincent Y. F. Tan, Song Bai  
**arXiv:** [2306.14435](https://arxiv.org/abs/2306.14435)  
**Date:** 2023-06-26  
**Summary page:** https://www.summarizepaper.com/en/paper/2306.14435v1/

## Key points

- DragDiffusion is a framework for interactive point-based image editing
- It leverages diffusion models to achieve precise spatial control
- Image editing requires pixel-level precision and has gained significant attention in recent years
- Existing method DragGAN allows for interactive point-based image editing but its generality is limited by pre-trained GAN models
- DragDiffusion extends the editing framework to diffusion models and utilizes large-scale pre-trained diffusion models to improve applicability in real-world scenarios
- Unlike most existing diffusion-based image editing methods, DragDiffusion optimizes the diffusion latent to achieve precise spatial control, enabling high-quality editing efficiently and effectively while also being applicable in real-world scenarios.
- Optimizing the diffusion latent at one single step suffices to generate coherent results.
- The versatility and generality of DragDiffusion are demonstrated through extensive experiments across a wide range of challenging cases such as multi-objects, diverse object categories, various styles etc., significantly improving upon existing methods by providing greater flexibility and precision in image editing tasks.
- Authors include Yujun Shi, Chuhui Xue, Jiachun Pan, Wenqing Zhang, Vincent Y.F Tan and Song Bai.

## Summary

DragDiffusion is a novel framework for interactive point-based image editing that leverages diffusion models to achieve precise spatial control. Image editing is a challenging task that requires pixel-level precision and has attracted significant attention in recent years. While DragGAN, an existing method, allows for interactive point-based image editing with impressive results, its generality is limited by the capacity of pre-trained GAN models. To overcome this limitation, DragDiffusion extends the editing framework to diffusion models and utilizes large-scale pre-trained diffusion models to improve the applicability of interactive point-based editing in real-world scenarios. Unlike most existing diffusion-based image editing methods that work on text embeddings, DragDiffusion optimizes the diffusion latent to achieve precise spatial control. This enables high-quality editing efficiently and effectively while also being applicable in real world scenarios. Although diffusion models generate images iteratively, DragDiffusion empirically shows that optimizing the diffusion latent at one single step suffices to generate coherent results. The versatility and generality of DragDiffusion are demonstrated through extensive experiments across a wide range of challenging cases such as multi-objects, diverse object categories, various styles etc., significantly improving upon existing methods by providing greater flexibility and precision in image editing tasks. The authors of this work include Yujun Shi, Chuhui Xue, Jiachun Pan, Wenqing Zhang, Vincent Y.F Tan and Song Bai.