To make full use of the complementary information in infrared and low-light remote sensing images and make it more convenient for visual interpretation, an image fusion method based on variation is proposed. In this variation-based mode, both the detail injection term and the structure fidelity term are defined. The spatial detail and structure characteristics of the fused images are also improved while the spectral characteristics of the infrared and low-light images are kept. A regularity energy term is incorporated into the fusion model so as to ensure the smoothness of the solution. On the basis of gradient descent flow, the fused images are obtained by numerical iteration. The experimental results show that the model can obtain the fused images containing abundant spatial and spectral information. Compared with the Laplacian pyramid decomposition-based and undecimated wavelet transform-based methods, the proposed model exhibits better fusion performance.