FEAD: Fine-Grained Epipolar Attention Diffusion for Large-Disparity Light Field Spatial Super-Resolution

Light-field (LF) spatial super-resolution hinges on exploiting cross-view correspondences to recover high-frequency details while preserving geometric structure. However, existing methods often rely on implicit feature interaction or coarse disparity maps, which suffers noticeable performance degradation in regions with large disparities. In this paper, we propose Fine-Grained Epipolar Attention Diffusion (FEAD), which adopts epipolar attention as an explicit mechanism to characterize cross-view geometric correspondence along epipolar lines. To enable more accurate feature alignment and long-range spatial–angular interaction, we introduce a diffusion-based epipolar attention refinement strategy that progressively improves the initial attention. With the refined attention as guidance, cross-view features are explicitly warped and fused to enforce geometric consistency and aggregate complementary details across views for fine-grained reconstruction. Extensive experiments demonstrate that FEAD achieves state-of-the-art performance, with particularly strong gains in large-disparity scenarios.