Author name: Caiyong Piao, Zhiyuan Yan, Haoming Xu, Yunzhen Zhao, Kaiqing Lin, Feiyang Xu, Shuigeng Zhou

arXiv:2603.01228v2 Announce Type: replace
Abstract: Accurate rejection of sensitive or harmful visual content, i.e., harmful image guardrail, is critical in many application scenarios. This task must continuously adapt to the evolving safety policies …