Co-occurring Associated REtained concepts in Diffusion Unlearning

[ICLR 2026 Accepted]

📄 Paper (OpenReview)

Updates

• 02-10-2026: Code released.

🔍 Framework Overview

This repository provides the official implementation of ReCARE, a diffusion model unlearning framework that removes a target concept while explicitly preserving benign co-occurring concepts (CARE).

Existing unlearning methods often suppress benign concepts that naturally co-occur with the erase target (e.g., erasing nudity unintentionally removes person).

ReCARE addresses this issue by:

• Automatically constructing a CARE-set from target images
• Integrating the CARE-set into unlearning training to preserve co-occurring semantics

Getting Started

🛠️ Environment Setup

git clone https://github.com/damilab/CARE.git
cd CARE

# Linux (recommended)
conda env create -f environment.yaml
conda activate recare

# clip
pip install git+https://github.com/openai/CLIP.git

🎨 Training Image Generation

Generate images containing the erase target. These images are later used both for CARE-set construction and unlearning training.

# Object Concept (Nudity)
python generate_images.py --output_dir data/nudity/images/ --prompt "A photo of a nude person" --num_images 500

# Style Concept (Van Gogh)
python generate_images.py --output_dir data/vangogh/images/ --prompt "A painting in the style of Van Gogh" --num_images 500

📚 CARE-set Construction

Construct the CARE-set, a curated vocabulary of benign co-occurring concepts.

# Object Concept (Nudity)
python utils/dictionary.py --reference_word 'nudity' --image_dir data/nudity/images/ --output_dir data/nudity/ --prompt_style photo

# Style Concept (Van Gogh)
python utils/dictionary.py --reference_word 'Van Gogh' --image_dir data/vangogh/images/ --output_dir data/vangogh/ --prompt_style style

This step produces a careset.json file, which serves as the CARE-set anchor during training.

🚀 ReCARE Training

# Object Concept (Nudity)
python -W ignore train.py --erase_concept 'nudity' --train_method noxattn --train_data_dir data/nudity/images/ --learnable_property 'object' --initializer_token 'person' --output_dir recare_weights/nudity --compositional_guidance_scale 2 --n_iterations 2 --num_of_adv_concepts 2 --anchor_concept_path data/nudity/careset.json

# Style Concept (Van Gogh)
python -W ignore train.py --erase_concept 'Van Gogh' --train_method noxattn --train_data_dir data/vangogh/images/ --learnable_property 'style' --initializer_token 'art' --output_dir recare_weights/vangogh --compositional_guidance_scale 2 --n_iterations 2 --num_of_adv_concepts 2 --anchor_concept_path data/vangogh/careset.json

📊 Evaluating CARE Preservation

Evaluate whether benign co-occurring concepts are preserved after unlearning, using the proposed CARE score.

python metrics/care_eval.py \
  --gen_prompts_file prompts_person.txt \
  --targets 'person' \
  --num_images 10 \
  --unet_checkpoint recare_weights/nudity/ReCARE-Diffusers-UNet.pt \
  --out_dir care_outputs \
  --save_images

Additional evaluation scripts (ASR, COCO-based FID, CLIP score) are available in metrics/.
For style-based evaluation, we use the pretrained style classifier released by Diffusion-MU-Attack. Please download the checkpoint and place it at: metrics/style_classifier/checkpoint-2800

💾 Pretrained Models

Pretrained ReCARE model checkpoints (ReCARE-Diffusers-UNet.pt) are provided via Google Drive. After downloading, place them under the recare_weights/ directory (e.g., recare_weights/nudity/).