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

TL;DR

FlowErase-RL introduces a reward-based framework for concept erasure in flow matching models, enhancing safety and control in text-to-image generation without requiring supervised data.

Contribution

It reformulates concept erasure as a reward optimization problem using a dual-path mechanism, enabling scalable and effective multi-concept erasure without explicit supervision.

Findings

Achieves state-of-the-art erasure performance on various concepts.

Maintains high image quality and semantic alignment after erasure.

Demonstrates robustness against adversarial attacks and scalability to multiple concepts.

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.