GEM: Geometric Erasure by Contrastive Velocity Matching in Rectified Flows
Researchers have introduced GEM (Geometric Erasure by Contrastive Velocity Matching in Rectified Flows), a framework for erasing unwanted concepts from generative models. This is particularly relevant as the field shifts towards Rectified Flow Transformers, which are more complex and harder to control than previous U-Net-based diffusion models. The authors claim that GEM can effectively suppress unwanted content while preserving benign generation. Their approach combines traj
Researchers have introduced GEM (Geometric Erasure by Contrastive Velocity Matching in Rectified Flows), a framework for erasing unwanted concepts from generative models. This is particularly relevant as the field shifts towards Rectified Flow Transformers, which are more complex and harder to control than previous U-Net-based diffusion models. The authors claim that GEM can effectively suppress unwanted content while preserving benign generation. Their approach combines trajectory-based unlearning with teacher-guided erasure, providing a unified framework for concept erasure.
---
Why it matters: This matters because generative models have the potential to create synthetic content that infringes on copyrights or spreads misinformation. By developing effective methods for erasing unwanted concepts, researchers can help mitigate these risks and ensure that AI-generated content is safe and responsible.
Source: https://arxiv.org/abs/2606.00140
This article was originally published at: https://arxiv.org/abs/2606.00140