Cyber violence severely disrupts public order in both cyberspace and the real world. Existing studies have gradually advocated collaborative governance but rely on macro-level theoretical analyses. This study integrates micro- and macro-level perspectives to propose a two-stage, multi-scenario governance mechanism for cyber violence. In the first phase, a multi-scenario evolutionary game model with four parties involved in cyber violence was developed based on evolutionary game theory. Matlab simulations show that under strong government regulation, moderate levels of punishment implemented by the government against the online media that adopt misguidance strategies can achieve the most desirable stable state. In the second phase, the role of bystanders was introduced by integrating communication dynamics theory, and emotional factors were considered alongside game strategies. This led to the development of a new SBI1I2R model for public opinion dissemination in cyber violence. Netlogo simulations found that increasing the "correct guidance" strategy by the online media reduces the influence of cyber violence supporters and the time it takes for their nodes to drop to zero, but does not significantly shorten the time for the peak to occur. Comparatively, collaborative intervention between the online media and the government was most effective in curbing public opinion, followed by the government's independent "strong regulation." Relying solely on the online media's "correct guidance" produced the weakest effect. Finally, this mechanism was applied to a case study, and a multi-stage, multi-scenario analysis based on life cycle theory enhanced its practical applicability.