Free-view image compression is thriving due to the widespread adoption of immersive visual systems in daily life. However, effectively utilizing the spatial relationships among multiple views in free-view image sequences poses a significant challenge, and existing learning models are primarily limited to stereo and light field image compression. In this work, we first propose an end-to-end compression network for free-view images, aiming to alleviate spatial redundancy among multiple views effectively. In our methods, we leverage contrastive learning loss to enable the codec to compress various views of the same scene with higher quality. Furthermore, a depth-based prediction module is designed to enhance prediction accuracy and conserve bit rate. Extensive experiments validate that our model can generate more satisfying reconstructed images and outperform the state-of-the-art models.