FlowErase-RL: Reframing concept erasure in flow matching models as reward optimization

FlowErase-RL reframes concept erasure in flow matching models as reward optimization, and the researchers claim it is the first such framework based on GRPO. Its dynamic dual-path reward uses a concept erasure reward to suppress the target concept and a non-target space reward to preserve generation fidelity, balanced by a performance-driven switching strategy, without explicit supervision. It achieves SOTA on nudity, object and artistic style erasure, maintains image quality and semantic alignment, resists adversarial attacks, and scales to multi-concept scenarios.