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

2026-04-17

AI Summary

Text-to-image generation models are susceptible to implicit biases in training data, often producing unsafe content, while existing concept erasure methods struggle to balance thoroughness with generation fidelity. This work proposes TICoE, a novel framework that introduces, for the first time, a text-image collaborative erasure mechanism. By integrating continuous convex concept manifold modeling with hierarchical visual representation learning, TICoE enables precise and minimally disruptive removal of target concepts. Furthermore, the authors devise a fidelity-oriented quantitative evaluation strategy. Extensive experiments demonstrate that TICoE significantly outperforms current approaches across multiple benchmarks, effectively eliminating undesirable content while preserving semantic coherence and visual quality of generated images, thereby enhancing model safety and controllability.

Technology Category

Computer Vision: Diffusion Models for VisionMachine Learning: Deep Generative Models & AutoencodersNatural Language Processing: Generation

Application Category

User Modeling, Personalization and Recommendation: Fairness-aware retrieval and rankingWeb Mining and Content Analysis: Large pretrained models with web dataSearch and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

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

Problem

• concept erasure
• text-to-image generation
• bias mitigation
• content safety
• semantic preservation

Innovation

• concept erasure
• text-to-image generation
• text-image collaboration
• convex concept manifold
• hierarchical visual representation