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

2026-05-19

AI Summary

This work addresses the safety risks of harmful content generation in flow-matching models by reframing concept erasure as a reward optimization problem. It introduces, for the first time, the GRPO reinforcement learning framework, equipped with a dynamic dual-path reward mechanism that jointly optimizes suppression of target concepts and fidelity of non-target generation, alongside a performance-driven adaptive switching strategy that enables stable training without explicit supervision. The proposed method achieves state-of-the-art results across multiple concept erasure tasks—including nudity, object removal, and artistic style elimination—significantly outperforming existing approaches. It demonstrates strong robustness and scalability while preserving high-quality image synthesis and semantic alignment.

Abstract

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.

Problem

• concept erasure
• flow matching models
• reward optimization
• text-to-image generation
• safety

Innovation

• Flow Matching
• Concept Erasure
• Reward Optimization
• Reinforcement Learning
• Dual-Path Reward