The paper introduces AdaRevD, an innovative approach to image deblurring that leverages the capabilities of a well-trained encoder and refactors a reversible decoder to enhance its decoding capability. AdaRevD consists of multiple sub-decoders, each trained in a reversible manner, which allows for efficient memory usage while maintaining high model capacity. The method disentangles high-level degradation degrees and low-level blur patterns from a compact degradation representation, enabling better deblurring performance. Additionally, a classifier is introduced to predict the degradation degree of image patches, allowing them to exit at different sub-decoders for speedup. Experiments on various datasets, including GoPro, HIDE, RealBlur-R, and RealBlur-J, demonstrate that AdaRevD achieves state-of-the-art results, achieving a PSNR of 34.60 dB on the GoPro dataset. The paper also discusses the effectiveness of the reversible architecture, the impact of the adaptive patch-exiting strategy, and the disentanglement of degradation representation and blur pattern.The paper introduces AdaRevD, an innovative approach to image deblurring that leverages the capabilities of a well-trained encoder and refactors a reversible decoder to enhance its decoding capability. AdaRevD consists of multiple sub-decoders, each trained in a reversible manner, which allows for efficient memory usage while maintaining high model capacity. The method disentangles high-level degradation degrees and low-level blur patterns from a compact degradation representation, enabling better deblurring performance. Additionally, a classifier is introduced to predict the degradation degree of image patches, allowing them to exit at different sub-decoders for speedup. Experiments on various datasets, including GoPro, HIDE, RealBlur-R, and RealBlur-J, demonstrate that AdaRevD achieves state-of-the-art results, achieving a PSNR of 34.60 dB on the GoPro dataset. The paper also discusses the effectiveness of the reversible architecture, the impact of the adaptive patch-exiting strategy, and the disentanglement of degradation representation and blur pattern.