TL;DR
An efficient aligned kernel network (AKNet) is proposed, which innovatively employs super-large convolution kernels to capture global receptive fields with minimal computational overhead, effectively mimicking the long-range dependency modeling of transformers.
The performance of deep learning-based dehazing frameworks inevitably degrades when applied to real-world scenarios, primarily due to the severe domain shift and diverse degradation types. A major bottleneck is that existing literature heavily relies on the raw pixel domain, thereby neglecting the distinct spectral characteristics of hazy observations and underutilizing the latent representation capacity of deep networks for high-quality image reconstruction. Concurrently, while attention-driven feature fusion has advanced image restoration, current formulations often suffer from inadequate dimension-wise feature interactions and coarse integration strategies. Furthermore, long-range contextual dependencies and global low-frequency illumination properties inherent in hazing processes remain largely unexploited. In this paper, we propose a dehazing framework named DKS-Net which fully utilizes the physics guiding features and extracting structural information in the spatial domain. Motivated by the realization that the learning of information at multiple scales and frequency bands are important for the deep networks, we introduce a Kernel Selective Feature Extraction Module(KSFE) to effectively captures structural patterns via large-kernel convolutions with dynamic selection capabilities and multi-scale semantic cues. With the above techniques, our method can show state-of-the-art(SOTA) performance on synthetic datasets and real-world datasets, achieving competitive performance in visual quality.
PixRestore is presented, a VAE-free pixel-space Diffusion Transformer (DiT) for UIR, where the diffusion backbone is trained entirely from scratch, without relying on T2I pretraining.
Image deraining remains a fundamental challenge in computer vision, as rain streaks severely degrade visual quality, obscure scene content, and hinder downstream applications such as autonomous driving and video surveillance. Existing methods often struggle to balance global scene understanding with precise rain removal, resulting in either residual artifacts or over-smoothed textures. To address this issue, we propose MPINet, a Multi-Stage Progressive Image Restoration Network for image deraining. MPINet integrates illumination-aware modeling with global context learning to improve deraining performance under diverse lighting conditions. Specifically, the illumination-aware module generates illumination maps to enhance robustness in scenes with varying brightness, while the UniMetaFormer-based core captures global semantic information through dynamic transformations and attention mechanisms, enabling more effective discrimination between rain streaks and underlying image structures. Built upon the multi-stage restoration paradigm of MPRNet, MPINet adopts a hierarchical progressive framework with patch-based processing and deep supervision across three stages, allowing efficient feature refinement with reasonable model complexity. Experimental results show that MPINet effectively removes rain streaks of varying densities while preserving fine textures. Across all evaluated datasets, MPINet outperforms MPRNet by about 6.5% in PSNR and 1.3% in SSIM on average.