The progress of the automatic driving system, namely ADS, requires rigorous testing in various safety-critical scenarios to ensure its reliability in a real environment. However, most of the existing scenario generation methods can’t balance multiple objectives, such as safety, reality, and diversity, and improve learning, namely RL utility, which limits their effectiveness in revealing system vulnerabilities. In order to deal with these problems, DriVE-Net is proposed, which is a brand-new multi-objective framework for generating diverse, realistic, and danger-aware driving. DriVE-Net integrates four core dimensions: one is safety, which is quantified by advanced risk indicators, such as ADI and TIS; the other is reality, which is guaranteed by aligning the data distribution of the scenario with the real world; the third is diversity, which is achieved by targeted parameter space sampling; and the fourth is learning effectiveness, which is optimized by the gradual difficulty level of RL agents. The framework uses a dynamic weighting mechanism to adaptively balance these goals during training. The experimental results on Waymo open data sets show that DriVE-Net is superior to the baseline method, achieving a scenario collision rate of 50% to 80%, namely SCR, while maintaining a high trajectory accuracy with an average displacement error of 0.32 m. The ablation research highlights the key role of speed-based measurement standards in accurately detecting hazards. The generated scenario effectively improves the robustness of the reinforcement learning strategy by exposing rare edge cases, and its effectiveness is quantitatively verified by key safety trigger rate and other indicators. This work provides a multi-functional unified framework for high-quality scenario generation, and solves the long-standing gap in the automatic driving system, namely ADS safety critical test. Its application includes self-driving vehicle verification, policy training, and risk assessment, which has academic significance and industrial value. Future research directions include extending the framework.