Title: Geometric Erasure by Contrastive Velocity Matching in Rectified Flows

Original: arXiv:2606.00140v1 Announce Type: cross Abstract: While the rapid adoption of multimodal generative models offers immense potential, it has also increased the risks of harmful content synthesis, deepfakes, and copyright infringements. To address these challenges, concept erasure has emerged as a prospective safeguard. However, as the field gradually transitions from U-Net-based diffusion models to Rectified Flow Transformers, erasure research has struggled to keep pace. In this work, we introduce GEM, a simple but highly effective erasure framework for Rectified Flow models. As part of our contribution, we establish a principled bridge between trajectory-based unlearning grounded in Generative Flow Networks and classic teacher-guided erasure: we translate trajectory-based signals into a teacher-guided flow-matching setup that unifies the strengths of both paradigms. Concretely, a teacher provides complementary attraction and repulsion signals that we combine into a single geometric guidance objective, yielding targeted suppression of unwanted concepts while preserving benign generation.

Rewrite: arXiv:2606.00140v1 Announce Type: cross Abstract: The swift proliferation of multimodal generative models presents significant opportunities, yet it simultaneously amplifies threats such as the creation of malicious content, deepfakes, and violations of intellectual property rights. Concept erasure has recently been proposed as a viable mitigation strategy to counter these dangers. Nevertheless, as the industry shifts from traditional U-Net-based diffusion architectures to Rectified Flow Transformers, advancements in erasure techniques have lagged behind. This paper presents GEM, a straightforward yet powerful erasure framework specifically designed for Rectified Flow models. Our key contribution lies in forging a theoretical connection between trajectory-based unlearning—rooted in Generative Flow Networks—and conventional teacher-guided erasure methods. By converting trajectory-based cues into a teacher-guided flow-matching environment, we merge the advantages of both approaches. Specifically, a teacher model generates opposing attraction and repulsion signals, which are integrated into a unified geometric guidance objective. This mechanism effectively suppresses designated undesirable concepts without compromising the quality of safe, benign outputs.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC