The assumption that many forms of high-dimensional data, such as images, actually live on low-dimensional manifolds, sometimes known as the manifold hypothesis, underlies much of our intuition for how and why deep learning works. Despite the central role that they play in our intuition, data manifolds are surprisingly hard to measure in the case of high-dimensional, sparsely sampled image datasets. This is particularly frustrating since the capability to measure data manifolds would provide a revealing window into the inner workings and dynamics of deep learning models. Motivated by this, we introduce neural frames, a novel and easy to use tool inspired by the notion of a frame from differential geometry. Neural frames can be used to explore the local neighborhoods of data manifolds as they pass through the hidden layers of neural networks even when one only has a single datapoint available. We present a mathematical framework for neural frames and explore some of their properties. We then use them to make a range of observations about how modern model architectures and training routines, such as heavy augmentation and adversarial training, affect the local behavior of a model.