Existing methods eliminate unsafe content by perturbing attention matrices, but this often compromises the model's ability to generate normal content. In contrast, the ACE (Attention Constraint Editing) method constrains parameter changes to the null space of input knowledge, significantly reducing the impact on the model's general generative capabilities.
The code base is based on the diffusers package. To get started:
git clone https://github.com/rohitgandikota/unified-concept-editing.git](https://github.com/littlelittlenine/ACE-zero.git
cd ACE-zero
mkdir models
pip install -r requirements.txt
python /train-scripts/erase_mass_alphaedit.py --model_save_path /models/edit.pt --concepts_save_path /models/edit.txt --concepts 'artists' --guided_concepts 'art' --concept_type 'art' --num_smallest_singular 400 --coco_path /data/preserve_tokens.csv --lamda 100 --device 0
To moderate concepts (e.g. "violence, nudity, harm")
python /train-scripts/erase_nudity.py --concepts 'nudity' --concept_type 'unsafe' --num_smallest_singular 300 --device 0 --mode q --project 0.16
To debias concepts (e.g. "Doctor, Nurse, Carpenter") against attributes (e.g. "Male, Female")
python /mlx_devbox/users/wangruipeng/playground/paper_run/uce_nullspace/unified-concept-editing-main/rich_article/train_debias_nullspace.py --concepts 'professions5' --concept_type 'bias_profession' --model_save_path /models/debias_10.pt --concepts_save_path /models/debias_10.txt --lamda 10 --device 0 --coco_path /data/extracted_subjects_big_only_1000.csv --add_prompts True