🌟 🌟 ECCV 2024 | Arxiv | 🤗HuggingFace 🌟 🌟
Authors
Chao Gong*, Kai Chen*, Zhipeng Wei, Jingjing Chen, Yu-Gang Jiang
Fudan University
The code that has been preliminarily organized has been released.
-
Run
pip install -r requirements.txt
to install the required packages. -
You can check
scripts/
for running scripts.
The edited models of RECE can be found 🤗here.
For all concepts, the coefficients of Eq.3 are:
The regularization coefficients
- Nudity and unsafe concepts(I2P concepts),
$\lambda=1e-1$ . - Artistic styles,
$\lambda=1e-3$ . - Difficult objects (e.g., church and garbage truck),
$\lambda=1e-3$ . - Easy objects (e.g., English Springer, golf ball and parachute),
$\lambda=1e-1$ . - For other objects where erasing accuracies reach 0 using UCE, RECE's further erasure is not applied.
We will update the Arxiv version to correct/align the experiment settings.
If you find our work helpful, please leave us a star and cite our paper.
@article{gong2024reliable,
title={Reliable and Efficient Concept Erasure of Text-to-Image Diffusion Models},
author={Gong, Chao and Chen, Kai and Wei, Zhipeng and Chen, Jingjing and Jiang, Yu-Gang},
journal={arXiv preprint arXiv:2407.12383},
year={2024}
}
Some code is borrowed from UCE.