4D cine MRI is essential for assessing cardiac function, providing continuous observation of ventricular dynamics. However, its low through-plane resolution compromises anatomical continuity between slices and reduces the accuracy of key clinical metrics such as ventricular boundary delineation and ejection fraction (EF) estimation. We propose a Long-Axis guided Diffusion Autoencoder that integrates LAX-derived structural cues into an unsupervised through-plane interpolation framework. By leveraging the generative prior of diffusion modeling, our method enhances structural consistency and inter-slice continuity compared with existing models, achieving higher Dice and lower HD95 scores, with visualizations confirming chamber-level coherence. This approach offers a practical solution to the inherent resolution limitations of cine MRI and demonstrates strong potential for future volumetric and functional cardiac analyses.