Geometric Erasure by Contrastive Velocity Matching in Rectified Flows

Researchers introduce GEM, a concept erasure framework designed for Rectified Flow models that addresses the limitations of existing erasure techniques built for older U-Net diffusion architectures. The method combines trajectory-based unlearning with teacher-guided flow matching to suppress unwanted concepts in generative AI while preserving legitimate generation capabilities.

The transition from U-Net-based diffusion models to Rectified Flow Transformers represents a significant architectural shift in generative AI, but safety research has lagged behind implementation. GEM addresses this gap by establishing a principled connection between two previously separate approaches: trajectory-based unlearning from Generative Flow Networks and traditional teacher-guided erasure methods. This convergence is technically important because it leverages geometric guidance signals—attraction and repulsion vectors from a teacher model—to create targeted suppression of harmful concepts.

The broader context involves escalating concerns about multimodal generative models enabling harmful applications including deepfakes, copyright violations, and malicious content synthesis. As these models become more capable and widely deployed, defense mechanisms have become critical infrastructure. GEM's contribution extends erasure research into newer transformer-based architectures that increasingly replace diffusion models, ensuring safety measures keep pace with model evolution.

For developers and organizations deploying Rectified Flow models, this framework provides a concrete method to reduce harmful concept synthesis without wholesale model retraining. The geometric approach offers interpretability advantages, allowing practitioners to understand which concepts receive suppression signals. This matters for compliance efforts and responsible AI development across industries relying on generative models.

Future research should validate GEM's effectiveness against adversarial attempts to circumvent erasure and measure potential degradation in legitimate generation quality across diverse use cases. The framework's scalability to larger models and its interaction with fine-tuning techniques warrant investigation.

• →GEM enables concept erasure in Rectified Flow Transformers, addressing the gap as the field moves beyond U-Net architectures
• →The framework unifies trajectory-based unlearning with teacher-guided erasure using geometric guidance signals for targeted concept suppression
• →Suppresses unwanted concepts while maintaining benign generation quality through complementary attraction and repulsion signals
• →Provides practical safety mechanisms for preventing deepfakes and copyright infringement in modern generative AI systems
• →Establishes principled theoretical foundation connecting Generative Flow Networks to teacher-guided flow matching paradigms