Deep learning has gained rising prominence in the stochastic reconstruction of 3D porous media. However, most existing methods struggle to realize multiscale reconstruction since the difficulty in generating adequate voxels to simultaneously maintain representative macroscopic size and microscale pore features. To achieve efficient generation of large-sized porous media with preservation of multiscale structural characteristics, we propose a parallel reconstruction approach built upon the novel transdimensional generative networks (TGNet). This model consists of three generators that sequentially transform Gaussian noise into 1D to 3D tensor spaces corresponding to the edges, surfaces, and volumes of all structural units. The generated edges and surfaces serve as conditional inputs to subsequent generators, providing boundary constraints for higher-dimensional structures and thereby ensuring continuity between adjacent units during parallel reconstruction. The Wasserstein distance and residuals of boundary conditions are considered to maintain the morphological accuracy of reconstructed structures. The result of reconstruction experiments demonstrates that the TGNet takes only 6 minutes to generate large porous media of 2200³ voxels with high-fidelity pore structure and achieves over a 98% reduction in time cost compared to the existing methods. Using this model, the multiscale porous media with pore/throat radii spanning over three orders of magnitude can be efficiently reconstructed through spatial superposition of structures generated at different scales. By overcoming the long-standing trade-off between reconstruction size and resolution, TGNet can be regarded as a powerful tool for generating precise datasets to study the structure-property relationships of multiscale porous media