<p>Detecting anomalies in surveillance videos is inherently challenging due to the limited availability of labeled data and the ambiguous nature of abnormal events. To address these constraints, the proposed work presents a generative framework that simulates realistic training scenarios by predicting future frames based on the continuous flow of video sequences, thereby enhancing the detection of subtle and previously unseen events. The proposed approach integrates optical flow estimation using FlowNet to capture fine-grained motion dynamics, followed by adaptive frame sampling that prioritizes frames with significant motion variations. A bidirectional Generative Adversarial Network (GAN) architecture is used, consisting of forward and backward generators to predict future and past frames respectively, and a discriminator that distinguishes between real and synthesized frames. The model is trained using a combined loss function comprising adversarial and reconstruction losses to ensure both realism and temporal accuracy in the generated outputs. Anomalies are detected by evaluating the discrepancy between the predicted and actual frames, with higher reconstruction errors signaling potential abnormal behavior. This generative modeling strategy enables the system to learn complex temporal dependencies and subtle motion cues, enabling the detection of rare or previously unseen events. By leveraging simulated sequences during training, the framework overcomes the limitations of real-world datasets and enhances its generalization capability across diverse surveillance scenarios. The performance of the proposed approach is demonstrated through extensive evaluations on standard benchmark datasets, where it achieves competitive performance, highlighting its robustness and adaptability across diverse surveillance environments. The experimental evaluation demonstrates that the proposed generative framework attains competitive performance across multiple benchmark datasets. In particular, it achieves an AUC of 79.4% on UCSD Ped1 and 93.7% on UCSD Ped2, outperforming conventional autoencoder-based approaches and perception-driven GAN models. On more complex datasets, the method also shows strong generalization, obtaining 90.2% on CUHK Avenue and 92.6% on UCF Crime. These results highlight the model’s effectiveness in capturing fine-grained motion and appearance variations, as well as its robustness in addressing diverse real-world anomaly detection scenarios.</p>