Beyond Text Prompts: Precise Concept Erasure through Text-Image Collaboration

Text-to-image generative models have achieved impressive fidelity and diversity, but can inadvertently produce unsafe or undesirable content due to implicit biases embedded in large-scale training datasets. Existing concept erasure methods, whether text-only or image-assisted, face trade-offs: textual approaches often fail to fully suppress concepts, while naive image-guided methods risk over-erasing unrelated content. We propose TICoE, a text-image Collaborative Erasing framework that achieves precise and faithful concept removal through a continuous convex concept manifold and hierarchical visual representation learning. TICoE precisely removes target concepts while preserving unrelated semantic and visual content. To objectively assess the quality of erasure, we further introduce a fidelity-oriented evaluation strategy that measures post-erasure usability. Experiments on multiple benchmarks show that TICoE surpasses prior methods in concept removal precision and content fidelity, enabling safer, more controllable text-to-image generation. Our code is available at https://github.com/OpenAscent-L/TICoE.git

Key Contributions

TICoE framework using text-image collaboration with continuous convex concept manifold for precise concept removal
Hierarchical visual representation learning that preserves unrelated semantic and visual content during erasure
Fidelity-oriented evaluation strategy measuring post-erasure model usability and content preservation

🛡️ Threat Analysis

Data Poisoning Attack

The paper addresses concept erasure to mitigate implicit biases and unsafe content embedded in training data of generative models. This is a defense against data-level contamination that causes models to generate undesirable outputs. The threat model involves corrupted/biased training data leading to unsafe generation.