LFTramba: Comprehensive Information Learning for Light Field Image Super-Resolution via A Hybrid Transformer-Mamba Framework

Haosong Liu, Xiancheng Zhu, Huanqiang Zeng, Jianqing Zhu, Yifan Shi, Jing Chen, Junhui Hou; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2025, pp. 1137-1147

Comprehensive exploration of information in complex 4D Light Fields (LF) is crucial for advancing Light Field image Super-Resolution (LFSR). Existing state-of-the-art methods based on CNN or Transformer typically decompose the 4D LF into 2D subspaces, limiting feature learning to specific aspects such as disparity cues or spatial-angular correlations. Therefore, they exhibit deficiencies in comprehensive subspace information integration and insufficient cross-domain feature interaction. Moreover, the potential of hybrid architectures that leverage multiple computational paradigms remains largely unexplored in LFSR. To overcome these limitations, we propose LFTramba, a novel hybrid network architecture that facilitates collaborative learning across multiple subspaces through a Transformer-Mamba framework. Specifically, the Spatial-Angular Mamba Block (SAMB), built on an efficient visual state space representation, alternates between capturing complementary spatial-angular features in the spatial and angular domains. Concurrently, the Epipolar Plane Transformer Block (EPTB) extracts disparity features alternately from horizontal and vertical Epipolar Plane Image (EPI) domains. Extensive experiments on real and synthetic datasets confirm that LFTramba surpasses existing LFSR methods while maintaining high computational efficiency. Notably, it achieved 3rd in Track 1 and 5th in Track 2 of the NTIRE 2025 Light Field Image Super-Resolution Challenge, highlighting its superior performance.