DiaGRL is a prototype extension of GRL for image restoration. The goal is to investigate whether adding diagonal stripe attention can improve super-resolution performance on images with strong diagonal structures, such as buildings, roads, and repetitive architectural patterns in Urban100.
In the original GRL, anchored stripe attention is applied along horizontal and vertical directions. In DiaGRL, we replace some H/W stripe layers with diagonal stripe layers, denoted as D45 and D135. For these diagonal layers, the stripe-branch feature is rotated by ±45° before entering the existing Anchored Stripe Attention module. After attention computation, the output is inversely rotated back and concatenated with the window-attention branch.
We evaluated several stage-wise scheduling strategies on Urban100 X4 using the pretrained GRL-Tiny x4 checkpoint. The compared settings were the original GRL baseline, Schedule A with diagonal layers uniformly inserted across all stages, Schedule B with alternating H/W and diagonal stages, and Schedule C with diagonal attention concentrated in later stages.
Among the tested schedules, Schedule C showed the most stable improvement. It improved the average PSNR_Y by +0.0334 dB and SSIM_Y by +0.00037 over the GRL baseline. Image-wise analysis showed that Schedule C improved PSNR_Y on 79 out of 100 Urban100 images, with the largest gain of +0.4949 dB on img_045_SRF_4.
These results suggest that diagonal attention is more effective when applied mainly in later stages rather than uniformly across all stages. The current rotation-based implementation is a practical prototype, but it may introduce interpolation loss during rotation and inverse rotation. A future direction is to replace rotation-based alignment with direct diagonal indexing to avoid interpolation artifacts.
