Video anomaly detection (VAD) is a challenging task, as it involves detecting anomalies at the frame level in a video stream with video-level supervision. The majority of existing weakly-supervised VAD methodologies, which employ Multiple-Instance Learning (MIL) along with a top- ranking loss, often exhibit high false alarm rates for normal frames within abnormal videos. This pitfall comes from two main factors: the limited discriminatory capacity of the top- ranking loss and the inefficiency of MIL in aggregating information from normal video segments. In contrast, semi-supervised VAD approaches that rely solely on normal video data during training often enhance the prediction of normal frames within abnormal videos, but at the expense of accuracy in detecting abnormal frames. To enhance the advantages of both approaches, we propose a two-stage method, namely Normalizing Flow-Debiased Contrastive Learning (NF-DCL). In the first stage, a Normalizing Flow model is employed to learn the distribution of normal feature representations and generate diverse synthetic normal samples. In the second stage, these synthesized features guide a Debiased Contrastive Loss integrated with MIL and top- ranking loss, enhancing feature discrimination while alleviating label ambiguity and biased learning. Extensive experiments on UCF-Crime, ShanghaiTech, and XD-Violence demonstrate that NF-DCL consistently improves performance and significantly reduces false alarms, outperforming existing state-of-the-art methods.