Since multispectral images (MSIs) and RGB images (RGBs) have significantly different definitions and severely imbalanced information entropies, the spectrum transformation between them, especially reconstructing MSIs from RGBs, is a big challenge. We propose a new approach, the Taiji Generative Neural Network (TaijiGNN), to address the above-mentioned problems. TaijiGNN consists of two generators, G_MSI, and G_RGB. These two generators establish two cycles by connecting one generator’s output with the other’s input. One cycle translates the RGBs into the MSIs and converts the MSIs back to the RGBs. The other cycle does the reverse. The cycles can turn the problem of comparing two different domain images into comparing the same domain images. In the same domain, there are neither different domain definition problems nor severely underconstrained challenges, such as reconstructing MSIs from RGBs. Moreover, according to several investigations and validations, we effectively designed a multilayer perceptron neural network (MLP) to substitute the convolutional neural network (CNN) when implementing the generators to make them simple and high performance. Furthermore, we cut off the two traditional CycleGAN’s identity losses to fit the spectral image translation. We also added two consistent losses of comparing paired images to improve the two generators’ training effectiveness. In addition, during the training process, similar to the ancient Chinese philosophy Taiji’s polarity Yang and polarity Yin, the two generators update their neural network parameters by interacting with and complementing each other until they all converge and the system reaches a dynamic balance. Furthermore, several qualitative and quantitative experiments were conducted on the two classical datasets, CAVE and ICVL, to evaluate the performance of our proposed approach. Promising results were obtained with a well-designed simplistic MLP requiring a minimal amount of training data. Specifically, in the CAVE dataset, to achieve comparable state-of-the-art results, we only need half of the dataset for training; for the ICVL dataset, we used only one-fifth of the dataset to train the model, but obtained state-of-the-art results.