Data augmentation is a critical component in building modern deep learning systems. In this paper, we propose MFG-Augment, a novel data augmentation method based on the mean-field game theory. The central idea is to consider every image as a distribution over their pixel or feature space. By using Mean-field Game (MFG) theory, we can generate a time-continuous “path” from one distribution to another so that the points along the “path” are augmented images or features. The synthetic data can be used to improve the training of machine learning models, including deep neural networks. We demonstrate that the proposed technique has three main advantages: 1) It is a general data augmentation approach and can work with any data modality. 2) It can be applied in both label-variant transformation and label-agnostic transformation, which improve the affinity and diversity of the augmented data, respectively. 3) It can be utilized to generate both augmented images' pixels and features. 4) The transformation in the MFG-Augment is time-continuous and can theoretically generate an infinite number of augmented data.