Decreasing magnetic resonance (MR) image acquisition times can potentially make MR examinations more accessible. Prior arts including the deep learning models have been devoted to solving the problem of long MRI imaging time. Recently, deep generative models have exhibited great potentials in algorithm robustness and usage flexibility. Nevertheless, no existing such schemes that can be learned or employed directly to the k-space measurement. Furthermore, how do the deep generative models work well in hybrid domain is also worth to be investigated. In this work, by taking advantage of the deep en-ergy-based models, we propose a k-space and image domain collaborative generative model to comprehensively estimate the MR data from under-sampled measurement. Experimental comparisons with the state-of-the-arts demonstrated that the proposed hybrid method has less error in reconstruction and is more stable under different acceleration factors.