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

TL;DR

TICoE introduces a collaborative framework for precise concept erasure in text-to-image models, improving safety and control by effectively removing unwanted concepts while preserving content fidelity.

Contribution

It proposes a novel text-image collaborative erasing method using a convex concept manifold and hierarchical visual learning, surpassing prior techniques in accuracy and content preservation.

Findings

TICoE outperforms previous methods in concept removal precision.

The framework maintains high content fidelity after erasure.

Experiments demonstrate safer and more controllable image generation.

Abstract

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