Modulating Transformers With Explicit Degradation Information for Real-World Light Field Image Super-Resolution

Transformers have achieved promising performance in light field (LF) image super-resolution (SR). However, existing Transformer-based methods are typically developed under simplified degradation assumptions (e.g., bicubic degradation), and suffer from limited generalization capability to real-world LF images with diverse degradations. In this paper, we propose Degradation-guided Modulator (DeMo), which incorporates explicit degradation information into Transformer-based LF image SR. Specifically, DeMo formulates the degradation information as modulation factors and dynamically modulates the generation of Query, Key, and Value to prevent complex degradations from misguiding the attention calculation. Extensive experiments on both synthetic and real LF datasets demonstrate that integrating DeMo into existing Transformer baselines consistently improves SR performance on real LF images, while maintaining their remarkable robustness to large-disparity variations.