top of page

latest stuff in ai, directly in your inbox. 🤗

Thanks for submitting!

DragonDiffusion: Enabling Drag-style Manipulation on Diffusion Models

Artificial Intelligence (AI) has undeniably revolutionized how we perceive and interact with digital imagery. A growing area in this field is the conversion of text to image, which is the basis of Text-to-Image (T2I) models. While existing T2I models possess the ability to produce high-quality images from intricate textual descriptions, they often fall short when it comes to precise editing of the generated or real images. Addressing this gap, Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang present DragonDiffusion, a groundbreaking image editing method.

DragonDiffusion offers a pioneering approach to enabling drag-style manipulation on diffusion models. What makes this method novel is its construction of classifier guidance that leverages the strong correspondence of intermediate features within the diffusion model. This approach empowers the editing signals to transform into gradients via feature correspondence loss, subsequently modifying the intermediate representation of the diffusion model.

In order to encompass both semantic and geometric alignment, the researchers built upon this guidance strategy and introduced a multi-scale guidance system. This results in a more holistic and encompassing guidance strategy that doesn't merely focus on one facet of image manipulation but instead brings together a range of perspectives.

In addition, the authors introduced a cross-branch self-attention mechanism to maintain consistency between the original image and the editing result. This is crucial as it ensures the original context and features of the image are preserved and respected during the editing process.

DragonDiffusion, with its efficient design, enables various editing modes for both generated and real images. This includes moving objects within the image, resizing objects, replacing the appearance of objects, and content dragging. Impressively, the team has ensured that all editing and content preservation signals originate from the image itself. This eliminates the need for model fine-tuning or additional modules, streamlining the editing process and making it more user-friendly.

DragonDiffusion is a significant stride forward in the realm of AI-assisted image editing. It enhances the capacity for precise, intricate editing without compromising the quality of the original image. This offers exciting potential for professionals in industries ranging from graphic design and advertising to technology and media.

To facilitate further development and usage, the source code for DragonDiffusion is available at The creators invite AI enthusiasts, developers, and researchers to delve into this innovative tool, further exploring the boundaries of AI-assisted image editing and manipulation.

31 views0 comments



Snapy allows you to edit your videos with the power of ai. Save at least 30 minutes of editing time for a typical 5-10 minute long video.

- Trim silent parts of your videos
- Make your content more interesting for your audience
- Focus on making more quality content, we will take care of the editing

Landing AI

A platform to create and deploy custom computer vision projects.


An image enhancement platform.


A tool for face-morphing and memes.


SuperAGI is an open-source platform providing infrastructure to build autonomous AI agents.


A tool to create personalized fitness plans.


A tool to summarize lectures and educational materials.


A platform for emails productivity.


An all-in-one social media management tool.


A tool to generate personalized content.

Addy AI

A Google Chrome Exntesion as an email assistant.


A telegrambot to organize notes in Notion.

bottom of page