Bird’s-Eye-View (BEV) perception has emerged as a cornerstone of multi-camera scene understanding in autonomous driving. However, its reliability degrades significantly when camera views are partially lost or corrupted, leading to incomplete spatial semantics and poor handling of dynamic objects. Existing generative methods alleviate view loss but remain limited to static scene recovery due to the lack of temporal reasoning. To address these challenges, we propose Temporal Diffusion-Masked BEV (TD-MBEV), a unified framework that integrates two complementary components: a Diffusion-based BEV Completion (DBC) module for high-fidelity reconstruction of static features, and a Temporal Dynamic Prediction (TDP) module for motion-aware modeling of dynamic objects. By jointly leveraging generative diffusion and temporal continuity, TD-MBEV ensures both spatial completeness and temporal consistency under sensor failures. Comprehensive evaluations on the NuScenes benchmark show that TD-MBEV consistently outperforms state-of-the-art baselines, including BEVStereo and PETRv2. In scenarios where the rear camera is absent, TD-MBEV achieves up to 2.18% improvement in NDS and 1.10% in mAP over PETRv2. Ablation studies further validate the complementary strengths of combining spatial and temporal modeling, establishing TD-MBEV as a robust solution for 3D object detection in challenging real-world driving environments.