FlowErase-RL: Rethinking Concept Erasure as Reward Optimization in Flow Matching Models

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Sun, Yi, Zhang, Zhiqi, Zhong, Xinhao, Zhou, Yimin, Sun, Shuoyang, Chen, Bin, Xia, Shu-Tao, Xu, Ke
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866913161062383616
author Sun, Yi
Zhang, Zhiqi
Zhong, Xinhao
Zhou, Yimin
Sun, Shuoyang
Chen, Bin
Xia, Shu-Tao
Xu, Ke
author_facet Sun, Yi
Zhang, Zhiqi
Zhong, Xinhao
Zhou, Yimin
Sun, Shuoyang
Chen, Bin
Xia, Shu-Tao
Xu, Ke
contents Recent advances in flow matching models have significantly improved text-to-image generation quality, but also introduce growing safety risks due to the generation of harmful or undesirable content. Existing concept erasure methods are either inference-time interventions with limited effectiveness or rely on supervised fine-tuning (SFT), which requires precisely aligned data and struggles with scalability and multi-concept settings. In this paper, we propose \emph{FlowErase-RL}, the first GRPO-based framework for concept erasure in flow matching models. We reformulate concept erasure as a reward optimization problem and introduce a \textbf{dynamic dual-path reward mechanism} that jointly optimizes (i) a Concept Erasure (CE) reward to suppress target concepts and (ii) a Non-target Space (NS) reward to preserve generative fidelity. The two reward paths are adaptively balanced during training via a performance-driven switching strategy, enabling stable optimization without explicit supervision. Extensive experiments on nudity, object, and artistic style erasure demonstrate that our method achieves state-of-the-art erasure performance while maintaining strong image quality and semantic alignment. Moreover, it exhibits robust resistance to adversarial attacks and scales effectively to multi-concept scenarios. Our results establish a new paradigm for safe and controllable generation in flow matching models.
format Preprint
id arxiv_https___arxiv_org_abs_2605_19739
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle FlowErase-RL: Rethinking Concept Erasure as Reward Optimization in Flow Matching Models
Sun, Yi
Zhang, Zhiqi
Zhong, Xinhao
Zhou, Yimin
Sun, Shuoyang
Chen, Bin
Xia, Shu-Tao
Xu, Ke
Computer Vision and Pattern Recognition
Recent advances in flow matching models have significantly improved text-to-image generation quality, but also introduce growing safety risks due to the generation of harmful or undesirable content. Existing concept erasure methods are either inference-time interventions with limited effectiveness or rely on supervised fine-tuning (SFT), which requires precisely aligned data and struggles with scalability and multi-concept settings. In this paper, we propose \emph{FlowErase-RL}, the first GRPO-based framework for concept erasure in flow matching models. We reformulate concept erasure as a reward optimization problem and introduce a \textbf{dynamic dual-path reward mechanism} that jointly optimizes (i) a Concept Erasure (CE) reward to suppress target concepts and (ii) a Non-target Space (NS) reward to preserve generative fidelity. The two reward paths are adaptively balanced during training via a performance-driven switching strategy, enabling stable optimization without explicit supervision. Extensive experiments on nudity, object, and artistic style erasure demonstrate that our method achieves state-of-the-art erasure performance while maintaining strong image quality and semantic alignment. Moreover, it exhibits robust resistance to adversarial attacks and scales effectively to multi-concept scenarios. Our results establish a new paradigm for safe and controllable generation in flow matching models.
title FlowErase-RL: Rethinking Concept Erasure as Reward Optimization in Flow Matching Models
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2605.19739