<p>This paper addresses the challenges of low global information utilization efficiency and insufficient local detail recovery in dynamic scene image deblurring tasks by proposing a novel dynamic deblurring algorithm based on cross-scale feature collaboration, termed M2-GAN (Mamba2-Generative Adversarial Network). The algorithm features a global-local dual branch generator, a cross-scale discriminator, and a dynamic weighted loss function. The global-local dual-branch generator comprises a global branch and a local branch. The global branch, centered around Mamba2, incorporates a multi-scale feature extraction module designed specifically for this study, enabling efficient extraction of global contextual information from images. Conversely, the local branch employs a residual dense module developed in this research to enhance detail expression in blurred regions. The cross-scale attention fusion module within the cross-scale discriminator adaptively fuses discriminative features from both the global and local branches. Furthermore, the proposed dynamic weighted hybrid loss function integrates pixel-level loss, perceptual loss, and adversarial loss to guide the network in enhancing detail recovery while maintaining structural consistency. Experimental results validate the effectiveness and advancement of the proposed algorithm in dynamic scene image deblurring tasks, demonstrating its superiority over other comparative deblurring algorithms and its capability to effectively eliminate blur while preserving the integrity and naturalness of image details.</p>