Tight oil and gas reservoirs have attracted many attentions and are one of the hottest research fields in recent years. Tight sandstones have complex pore structures and narrow pores and throats with pore sizes varying from nanometers to micrometers, studying flow mechanisms in tight sandstones is of significance to tight oil/gas reservoir development. Reconstructing the digital rock which can comprehensively represent petrophysical properties of tight sandstone is the key to simulate the fluid flow in micro/nano pores. This paper proposes a new method of reconstructing 3D digital rock from CT image of tight sandstones based on a deep convolutional generative adversative network (DCGAN), and 3D convolution in the generator and discriminator are adopted to realize reconstruction from one dimensional data to 3D digital rock model. Studies show that when the training effect is slightly poor, the generated digital rock model will have noise, which can be reduced by post-processing; when the training effect is well, DCGAN can accurately reconstruct the 3D digital rock of tight sandstones, the reconstructed digital rock is very consistent in pore size, geometric structure, and connectivity of natural tight sandstones. When multiple 3D tight sandstone CT images are used for training, the DCGAN can learn the pore structure characteristics of entire tight sandstones, and the porosity distribution obtained from generated digital rock are similar to original tight sandstones.