SwapAnything: Enabling Arbitrary Object Swapping in Personalized Visual Editing (ECCV 2024)

Jing Gu, Yilin Wang, Nanxuan Zhao, Wei Xiong, Qing Liu, Zhifei Zhang, He Zhang, Jianming Zhang, HyunJoon Jung, Xin Eric Wang

[Project Page] [Paper]

Model Architecture

TODO

• Release demo
• Release code

Crop-Edit-Paste Pipeline

Since swap-anything could preserve the background perfectly, we could avoid the image size and ratio limitation of the backbone image diffusion model. Now we could edit object of any size in image of any size! Swap-anything support personalized object swapping, general object swapping, and object insertion.

1. Git clone this repo

git clone https://github.com/eric-ai-lab/swap-anything.git
cd swap-anything

2. Image and Mask Preparation

1. For a source image filename.jpg, please create source_image/filename/:

   mkdir source_image/filename
   

2. Put binary mask image into source_image/filename. The mask image name should be filename_mask.png.

We accept general source image types like .jpg, .jpeg, .png. However, make sure the mask is in .png format so that the code can find the mask according to pattern matching. Please refer to the provided example in source_image.

3. Train DreamBooth Model to Include the New Concept (Optional)

1. Our method also works with other concept learning methods such as CustomDiffusion, Text Inversion, etc.
2. You do not need this step for general object swapping.

Package Installation

Follow the installation instructions from Hugging Face Diffusers v0.25.0:

cd diffusers
pip install -e .
cd examples/dreambooth
pip install -r requirements.txt

Concept Data

Place images of the target concept into dreambooth_data/{INSTANCE_NAME}. The more images you provide, the better the results. Optimal performance is typically achieved with around 20 images.

Training DreamBooth

In script train_dreambooth.sh

1. Set INSTANCE_NAME to the name of the image folder.
2. Set CLASS_NAME to the class of the target object.

Run the following command to start training:

./train_dreambooth.sh

The script will generate checkpoints in the folder checkpoints/checkpoint_$INSTANCE_NAME-$Model_IDENTITY.

4. Editing Image

• Personal object swapping:

  python main.py --config config_personal_swap.yml
  

  To perform a personalized swap, set concept_model_path to the folder containing the diffusion model with the learned concept. Our model uses DreamBooth, so update source_subject_word with the object you want to replace and adjust source_prompt accordingly. Similarly, modify target_subject_word and target_prompt based on the tokens used during DreamBooth training.

• General object swapping:

  python main.py --config config_general_swap.yml
  

  For general object swaps (non-personalized), you can use "runwayml/stable-diffusion-v1-5" as the concept_model_path.

• Object insertion:

  python main.py --config config_insertion.yml
  

  For object insertion, follow the same process as the swapping task. The difference is to set source_subject_word to 'nothing' and source_prompt to "a photo of nothing".

Output

The editing results will be in the folder photoswap_real_output_{cuda_id}. In each sample_* subfolder, you will find the source image, edited image, mask, along with their cropped versions, and a JSON file containing all variable information.

We also provide a webpage in the html folder so you can browse all results at once.

config*.yml expanation

A Few Tips for Better Performance

1. For face swapping tasks, usually a higher variable swapping ratio yields better performance.
2. If the swapping results show significant shape deformation, it could be due to the automatic cropping issue. Manually input the crop coordinates x1, y1, x2, y2 so that (x2 - x1) ≈ (y2 - y1). This is also useful for detailed manipulation. For example:
   • Setting (x2 - x1) > (y2 - y1) will make the face wider.
   • Setting (x2 - x1) < (y2 - y1) will make the face narrower horizontally.
3. If the swapping result is too similar to the source image and does not transfer the target identity, try decreasing the swapping ratio. If the result contains artifacts or is not harmonious, you may want to increase the swapping ratio.

Citation

@inproceedings{gu2024swapanything,
      title={SwapAnything: Enabling Arbitrary Object Swapping in Personalized Visual Editing}, 
      author={Jing Gu and Yilin Wang and Nanxuan Zhao and Wei Xiong and Qing Liu and Zhifei Zhang and He Zhang and Jianming Zhang and HyunJoon Jung and Xin Eric Wang},
      booktitle={ECCV},
      year={2024}
}