Bulk Concepts Forgetting for Text-to-Image Diffusion Models

This is a project aiming to erase 10^3 concepts on T2I diffusion models while remaining its orginal performace, i.e., fidelity.

This project is not for scientific research purposes. JUST FOR FUN!!!

Inspried by MEMIT, we find that erasing concepts in text encoder is a much better way to achieve bulk concepts erasing.

Installation

conda create -n bcf python=3.10
conda activate bcf
cd bcf
pip install -r requirement.txt

Run BCF

• Forgetting Artistic Styles

python clip_edit.py --file 'data/art_forget/artist_list_forget_10.csv' --algorithm 'bcf' --concept_type 'artist_forget' --seed 42 --v_similarity_metric "cosine"

• Forgetting Objects

python clip_edit.py --file 'data/object_forget/object_erase_10.csv' --algorithm 'bcf' --concept_type 'object_forget' --seed 42 --v_similarity_metric "cosine"

• Forgetting NSFW

python clip_edit.py --file 'data/nsfw_forget/nsfw_list_forget_all.csv' --algorithm 'bcf' --concept_type 'NSFW' --seed 42 --v_similarity_metric "cosine"

Forgetting results are stored at images/results/bcf/<concept_type>/ .

Edited model are stored at model/<concept_type>/ .