Introducing MambaIR as a simple but effective baseline for image restoration, enhancing global receptive fields and computational efficiency.
The CU-Mamba model effectively captures both global spatial context and channel-wise feature correlations for high-quality image restoration, outperforming state-of-the-art methods while maintaining a lower computational cost.
This paper introduces a novel algorithm, PnP-Flow Matching, which leverages pre-trained Flow Matching models for image restoration tasks, achieving state-of-the-art results in denoising, deblurring, super-resolution, and inpainting by combining the strengths of Plug-and-Play methods and generative models.
HAIR, a novel hypernetwork-based approach to all-in-one image restoration, leverages dynamic parameter generation based on input image degradation information to outperform existing methods in both single-task and multi-task settings.
LoRA-IR, a novel framework for all-in-one image restoration, leverages low-rank experts and a CLIP-based degradation-guided router to achieve state-of-the-art performance and strong generalization across diverse image restoration tasks.