Ruibin Li¹ | Ruihuang Li¹ |Song Guo² | Lei Zhang^1*
1The Hong Kong Polytechnic University, ²The Hong Kong University of Science and Technology.
In ECCV2024

Pipelines of different inversion methods in text-driven editing. (a) DDIM inversion inverts a real image to a latent noise code, but the inverted noise code often results in large gap of reconstruction $D_{Rec}$ with higher CFG parameters. (b) NTI optimizes the null-text embedding to narrow the gap of reconstruction $D_{Rec}$, while NPI further optimizes the speed of NTI. (c) DirectInv records the differences between the inversion feature and the reconstruction feature, and merges them back to achieve high-quality reconstruction. (d) Our SPDInv aims to minimize the gap of noise $D_{Noi}$, instead of $D_{Rec}$, which can reduce the impact of source prompt on the editing process and thus reduce the artifacts and inconsistent details encountered by the previous methods.

## git clone this repository
git clone /~https://github.com/leeruibin/SPDInv.git
cd SPDInv

# create an environment with python >= 3.8
conda env create -f environment.yaml
conda activate SPDInv

python run_SPDInv_P2P.py --input xxx --source [source prompt] --target [target prompt] --blended_word "word1 word2"

python run_SPDInv_MasaCtrl.py --input xxx --source [source prompt] --target [target prompt]

To run PNP, you should first upgrade diffusers to 0.17.1 by

pip install diffusers==0.17.1

then, you can run

python run_SPDInv_PNP.py --input xxx --source [source prompt] --target [target prompt]

For ELITE, you should first download the pre-trained global_mapper.pt checkpoint provided by the ELITE, put it into the checkpoints folder.

python run_SPDInv_ELITE.py --input xxx --source [source prompt] --target [target prompt]

@article{li2024source,
  title={Source Prompt Disentangled Inversion for Boosting Image Editability with Diffusion Models},
  author={Li, Ruibin and Li, Ruihuang and Guo, Song and Zhang, Lei},
  booktitle={European Conference on Computer Vision},
  year={2024}
}

This code is built on diffusers version of Stable Diffusion.

Meanwhile, the code is heavily based on the Prompt-to-Prompt, Null-Text Inversion, MasaCtrl, ProxEdit, ELITE, Plug-and-Play, DirectInversion, thanks to all the contributors!.