This paper proposes VideoEraser, a novel framework that requires no training, to address privacy, copyright, and security concerns arising from exploitation of text-to-video (T2V) diffusion models. VideoEraser is designed as a plug-and-play module that can be integrated into existing T2V diffusion models through a two-step process: Selective Prompted Embedding Adjustment (SPEA) and Adversarial-Resilient Noise Guidance (ARNG). VideoEraser effectively prevents the creation of videos containing unwanted concepts such as objects, artistic styles, celebrities, and explicit content. Experimental results show that VideoEraser outperforms existing methods in efficiency, integrity, fidelity, robustness, and generalization performance, achieving an average of 46% unwanted content reduction across four tasks, achieving state-of-the-art performance.
