NullSCE: Sequential Concept Erasure in Generative Video Diffusion Models via Null-space Guidance

Text-to-video models (T2Vs) trained on large, unfiltered video corpora often learn and reproduce content including harmful concepts. Concept erasure techniques aim to remove such harmful concepts from pre-trained models while preserving overall generative capability. Although existing methods have achieved some success, they are all limited to batch-wise concept erasure. Applying them sequentially in streaming settings leads to two critical issues: erasure-forgetting phenomenon, where later erasures reactivate earlier ones, and cumulative drift, where repeated weight updates degrade overall generative capability. Thus, we introduce NullSCE, a null-space-guided sequential concept erasure framework for T2Vs. NullSCE reformulates the batch-wise concept erasure as a constrained optimization problem augmented with cumulative constraints that lock in previously erased concepts, and a norm-ball regularizer that bounds parameter drift. To solve it efficiently, we project each new erasure gradient onto the null space of prior conceptual erasure targets' gradients, ensuring orthogonality and minimizing inter-step interference. Through extensive experiments on state-of-the-art T2Vs and real-world datasets, we show that NullSCE reliably prevents concept reactivation, controls drift accumulation, and preserves high-fidelity video synthesis.