CRCE is a novel concept erasure framework for text-to-image diffusion models that handles coreferential concepts (synonyms, related terms) to prevent bypass attacks while preserving model utility.

• CRCE: Coreference Retention Concept Erasure in Text-to-Image Diffusion Models
  • Table of Contents
  • Method
  • Installation
  • Usage
  • Multi-Agent Framework
  • CorefConcept Dataset
  • Results
  • Citation
  • License

CRCE uses a multi-objective loss function:

L_total = L_anchor + α·L_coref + β·L_retain

• L_anchor: Erases target concept
• L_coref: Erases related concepts (synonyms, variations)
• L_retain: Preserves unrelated concepts

git clone https://github.com/vios-s/CRCE-Coreference-Retention-Concept-Erasure-in-Text-to-Image-Diffusion-Models
cd CRCE
pip install -r requirements.txt

from srcs.ours_tools import execute_ours_unlearn

# Erase "airplane" and related concepts while preserving other flying objects
result = execute_ours_unlearn(
    erase_concept="airplane",
    coref_concept="aeroplane,plane,jet plane,passenger plane",
    retain_concept="hot air balloon,blimp,rocket,drone",
    iterations=500,
    train_method='xattn-strict'
)

For automated experiments with LLM-guided concept identification:

python srcs/main.py  # Requires llmconfig.json with API keys

Pre-curated concept sets for reproducible experiments:

• CorefConcept/object.json - CIFAR-10 based objects
• CorefConcept/celebrity.json - Public figures
• CorefConcept/ip.json - Intellectual property

CRCE achieves 95%+ target concept removal while maintaining 85%+ retention quality, with 2-5x faster training compared to full model fine-tuning.

@inproceedings{xue2025crce,
  title={CRCE: Coreference Retention Concept Erasure in Text-to-Image Diffusion Models},
  author={Xue, Yuyang and Moroshko, Edward and Chen, Feng and Sun, Jingyu and McDonagh, Steven and Tsaftaris, Sotirios A},
  booktitle={British Machine Vision Conference (BMVC)},
  year={2025}
}

MIT License - see LICENSE file for details.