X-Restormer++: 1st Place Solution for the UG2+ CVPR 2026 All-Weather Restoration Challenge
Researchers have developed X-Restormer++, a novel deep learning model that won first place in the UG2+ Challenge at CVPR 2026 for all-weather image restoration. The model builds upon the X-Restormer architecture, incorporating dual-attention mechanisms and a spatially-adaptive input scaling method. A key innovation is the Gradient-Guided Edge-Aware (GGEA) Loss, which enhances the preservation of structural details by focusing supervision on edges and high-frequency regions. The solution employed a two-stage training strategy with a dual-model ensemble inference for optimal performance. AI
IMPACT Sets a new benchmark for all-weather image restoration, potentially improving applications in autonomous driving and surveillance.