Total generalized variation is effective and widely used in natural image denoising and deblurring due to its ability to suppress the staircase effect while preserving image edges and details. In order to improve the reconstruction performance of blind deconvolution on solar images, total generalized variation and PSF regularization are introduced into the reconstruction of solar images. A space-invariant multi-frame blind deconvolution model via second-order total generalized variation is proposed in this paper to improve the robustness of noise and recover more texture details. The model is solved by alternating minimization of the image sub-model and the PSF sub-model, where the image submodel can be solved by the half-quadratic splitting method. Combined with the non-blind deconvolution based on hyper-Laplacian prior, a space-invariant multi-frame blind deconvolution algorithm can be established under the multiscale framework. Then, by overlapping image segmentation and weighted stitching, the space-invariant blind deconvolution algorithm is extended to a reconstruction algorithm suitable for wide field-of-view solar images, which can reduce reconstruction errors caused by anisoplanatism. Finally, the reconstruction experiment and analysis are carried out on the real solar images observed by the one-meter New Vacuum Solar Telescope (NVST) in southwest China. The results show that the algorithm has good image reconstruction performance in both subjective visual effects and objective indexes. Second-order total generalized variation regularization and multi-frame can improve the stability and reliability of solar image reconstruction.Blind deconvolution is one of the commonly used post-reconstruction methods for adaptive optics images. In order to improve the reconstruction performance of blind deconvolution on solar (adaptive optics) images, a space-variant multi-frame blind deconvolution model based on second-order total generalized variation is proposed. It first solves the proposed space-invariant blind deconvolution model via second-order total generalized variation by the alternating minimization and half-quadratic splitting method. Then, according to the characteristics of wide field-of-view solar images which are anisoplanatic, the space-variant in the proposed algorithm is implemented by overlapping image segmentation and weighted stitching. Finally, the reconstruction experiment and analysis are carried out on the real solar images observed by the one-meter New Vacuum Solar Telescope (NVST). The results show that the proposed algorithm has good image reconstruction performance in both subjective visual effects and objective indexes.