Repository files navigation

This repository is the official implementation for the paper:Meta-Unlearning on Diffusion Models:Preventing Relearning Unlearned Concepts

We follow theDiffuser to install the required dependencies, please run the following commands:

conda create -n meta python=3.10conda activate metapip install --upgrade diffusers[torch]

All experiments are conducted on NVIDIA A100 GPUs with 80GB of memory.

First you can use the shell scripts to generatehrm dataset,irt dataset andtarget dataset. In our code, thehrm dataset contains the unlearned concept, while both theirt andtarget dataset contain retain concepts.irt dataset means the concept unrelated to unlearned concept, and thetarget dataset means the concept related to unlearned concept.

Here is the exmaple generate scipt, you can change the text prompt according to your goal.

bash scripts/gen_image.sh

bash scripts/esd_baseline.shbash scripts/sdd_baseline.sh

UCE andRECE can be trained by the original code in their paper.

For ESD and SDD based meta-unlearning:

bash scripts/esd_meta.shbash scripts/sdd_meta.sh

Note that we only give the example hyperparameter in code and you should change the hyperparameter refer to our paper (We usefix_timesteps = False in our paper, and thefixed_timesteps = True is just used for accelerating training) .

For UCE and RECE based meta-unlearning:

bash scripts/meta.sh

Note that you should change the model path to your unlearned model.

This repository is based on the codebase of theSDD. Thanks for their impressive works!

If you find our work helpful, please cite as

@article{gao2024meta,  title={Meta-Unlearning on Diffusion Models: Preventing Relearning Unlearned Concepts},  author={Gao, Hongcheng and Pang, Tianyu and Du, Chao and Hu, Taihang and Deng, Zhijie and Lin, Min},  journal={arXiv preprint arXiv:2410.12777},  year={2024}}