CRCE: Coreference-Retention Concept Erasure in Text-to-Image Diffusion Models

2025-03-18

AI Summary

Text-to-image diffusion models frequently generate harmful content, and existing concept erasure methods suffer from under-erasure (incomplete removal of target concepts) and over-erasure (unintended deletion of semantically or visually similar but irrelevant concepts). To address these issues, we propose CRCE—a novel Concept Rectification and Co-reference Erasure framework. CRCE is the first to integrate coreference relation modeling and semantic preservation mechanisms into concept erasure: it leverages large language models to identify co-referential concepts requiring joint erasure and discriminative semantics that must be retained. By combining concept-aware contrastive learning with targeted latent-space intervention in diffusion models, CRCE achieves precise, controllable concept removal. Extensive experiments across diverse erasure tasks demonstrate that CRCE significantly outperforms state-of-the-art methods—completely eliminating target concepts while preserving image fidelity and the integrity of unrelated concepts.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning
• Natural Language Processing: Generation

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Text-to-Image diffusion models can produce undesirable content that necessitates concept erasure techniques. However, existing methods struggle with under-erasure, leaving residual traces of targeted concepts, or over-erasure, mistakenly eliminating unrelated but visually similar concepts. To address these limitations, we introduce CRCE, a novel concept erasure framework that leverages Large Language Models to identify both semantically related concepts that should be erased alongside the target and distinct concepts that should be preserved. By explicitly modeling coreferential and retained concepts semantically, CRCE enables more precise concept removal, without unintended erasure. Experiments demonstrate that CRCE outperforms existing methods on diverse erasure tasks.

Problem

• Addresses under-erasure and over-erasure in text-to-image models
• Introduces CRCE for precise concept removal using Large Language Models
• Ensures retention of unrelated but visually similar concepts during erasure

Innovation

• Leverages Large Language Models for precise erasure
• Identifies related and distinct concepts semantically
• Prevents unintended erasure while removing target concepts