DesignEdit: Multi-Layered Latent Decomposition and Fusion for Unified & Accurate Image Editing

Stable Diffusion XL 1.0 Implementation

teaser

Project Page Paper Hugging Face Demo

✨ News ✨

[2024/4/4] We have supported the Gradio Application on Hugging Face 🤗, encouraging you to design online without the need for local deployment.
[2024/3/28] We release the code for DesignEdit! Let's design together! 😍

Setup

The required Python version is 3.10.12. , and the Pytorch version is 2.0.1. The code's framework is built on Prompt-to-prompt and Stable Diffusion.

Additional required packages are listed in the requirements file.

conda create -n DesignEdit python=3.10.12
conda activate DesignEdit
pip install -r requirements.txt

Notice that our model is entirely training-free💪!!! The base model is the Stable Diffusion XL-1.0.

Demo

We have created an interactive interface using Gradio, as shown below. You only need to simply run the following command in the environment we previously set up:

python design_app.py

page_1

🖱️Usage

We have 5 function pages for different editing operations.

💡Object Removal

💡Zooming Out

💡Camera Panning

💡Object Moving, Resizing and Flipping

💡Multi-Layered Editing

You can follow the "Usage" instructions within each page.

page_4

For each page, we also provide some interesting examples for you to try.

page_2

Notice that the Multi-Layered Editing page, which uses a multi-layered representation for multiple editing tasks, can achieve the same results as those of Object Removal and Object Moving, Resizing, and Flipping in a general representation.
Moreover, we have added the "Mask Preparation" page for you to utilize SAM or sketching to combine several masks together. This may be useful when you are on the Multi-Layered Editing page.

page_3

More Details

If you are interested in exploring more details about the model implementation, we recommend checking out model.py. Pay special attention to the register_attention_control() function and the LayerFusion class.

Applications

For more applications, we kindly invite you to explore our project page and refer to our paper.

💡Object Removal

You can choose more than one object to remove on the Object Removal page, and it is also possible to mask irregular regions for removal.

💡Object Removal with Refine Mask

Using remove mask directly may cause artifacts, the refine mask indicates regions that may cause artifacts. You can turn to Object Removal page to explore.

💡Camera Panning and Zooming Out

You can use the Camera Panning and Zooming Out page to achieve editing with different scales and directions.

The illustration of image adjustment and mask preparation is shown below.

💡Multi-Object Editing with Moving, Resizing, Flipping

You can achieve single object moving, resizing, flipping in Object Moving, Resizing and Flipping page, for multi-object editing like swapping and addition, you can turn to Multi-Layered Editing page.

💡Cross-Image Composition

By choosing one image as the background and specifying the position, size, and placement order of the foreground images, we can achieve cross-image composition. You can try examples on the Multi-Layered Editing page.

💡Typography Retyping

Typography retyping refers to the specific use of design elements, which you can achieve on the Multi-Layered Editing page.

Acknowledgements

Our project benefits from the contributions of several outstanding projects and techniques. We express our gratitude to:

Prompt-to-Prompt: For innovative approaches in prompt engineering.
Proximal-Guidance: For their cutting-edge inversion technique, significantly improving our model's performance.
DragonDiffusion: For inspiration on Gradio interface and efficient SAM API integration.

Each of these projects has played a crucial role in the development of our work. We thank their contributors for sharing their expertise and resources with the community.

BibTeX

@misc{jia2024designedit,
  title={DesignEdit: Multi-Layered Latent Decomposition and Fusion for Unified & Accurate Image Editing},
  author={Yueru Jia and Yuhui Yuan and Aosong Cheng and Chuke Wang and Ji Li and Huizhu Jia and Shanghang Zhang},
  year={2024},
  eprint={2403.14487},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}

design-edit / DesignEdit

readme