X-Restormer++ Wins CVPR 2026 All-Weather Restoration Challenge

A team of researchers has won the 8th UG2+ Challenge (CVPR 2026) Track 1 on Image Restoration under All-weather Conditions with their method X-Restormer++. The solution builds on the X-Restormer baseline, which uses a dual-attention design (Multi-DConv Head Transposed Attention and Overlapping Cross-Attention) to capture channel-wise global dependencies and spatially-local structural information. Key improvements include integrating a spatially-adaptive input scaling mechanism from Restormer-Plus to dynamically adjust spatial weights, introducing a novel Gradient-Guided Edge-Aware (GGEA) loss combined with L1 and Multi-Scale SSIM losses, and significantly expanding the training dataset. The work is detailed in a paper on arXiv (2605.13258).